Merge pull request 'feat(core): HTTP/2 support, correctness fixes, and doc reorganization' (#10) from feature/core/http2 into master

Reviewed-on: #10

.github/workflows/ci.yml

@@ -32,8 +32,40 @@ jobs:
           server-username: MAVEN_USERNAME
           server-password: MAVEN_PASSWORD
 
+      - name: Install h2spec 2.6.0
+        run: |
+          curl --fail --location --silent --show-error \
+            --output /tmp/h2spec.tar.gz \
+            https://github.com/summerwind/h2spec/releases/download/v2.6.0/h2spec_linux_amd64.tar.gz
+          echo "157ee0de702e01ad40e752dbf074b366027e550c8e7504f9450da2809e279318  /tmp/h2spec.tar.gz" \
+            | sha256sum --check
+          tar --extract --gzip --file /tmp/h2spec.tar.gz --directory /tmp
+
+      - name: Install nghttp client
+        run: |
+          sudo apt-get update
+          sudo apt-get install --yes nghttp2-client
+
+      - name: Install grpcurl 1.9.3
+        run: |
+          curl --fail --location --silent --show-error \
+            --output /tmp/grpcurl.tgz \
+            https://github.com/fullstorydev/grpcurl/releases/download/v1.9.3/grpcurl_1.9.3_linux_x86_64.tar.gz
+          echo "a926b62a85787ccf73ef8736b3ae554f1242e39d92bb8767a79d6dd23b11d1d5  /tmp/grpcurl.tgz" \
+            | sha256sum --check
+          tar --extract --gzip --file /tmp/grpcurl.tgz --directory /tmp grpcurl
+
       - name: Build and test
-        run: mvn -B --settings .github/settings.xml clean verify
+        run: >-
+          mvn -B --settings .github/settings.xml
+          -Dh2spec.executable=/tmp/h2spec
+          -Dcurl.executable=/usr/bin/curl
+          -Dnghttp.executable=/usr/bin/nghttp
+          -Dgrpcurl.executable=/tmp/grpcurl
+          -Djdk.tracePinnedThreads=full
+          -Pjmh
+          -Dflash.performance.gates=true
+          clean verify
         env:
           MAVEN_USERNAME: ${{ secrets.MAVEN_USERNAME }}
           MAVEN_PASSWORD: ${{ secrets.MAVEN_PASSWORD }}

README.md

@@ -1,12 +1,13 @@
 # Flash
 
-A high-performance HTTP/1.1 server library for Java 21, built around virtual threads and a zero-allocation FSM router.
+A high-performance HTTP/1.1 and HTTP/2 server library for Java 21, built around virtual threads,
+a zero-allocation FSM router, bounded protocol state, and one shared request/response API.
 
 ## Modules
 
 | Module | Description |
 |---|---|
-| `flash` | Core server library — router, request parser, HTTP I/O transport |
+| `flash` | Core server library — HTTP/1.1 and HTTP/2 transport, router, request/response model |
 | `flash-extensions/flash-ext-jackson` | Jackson JSON integration |
 | `flash-extensions/flash-ext-openapi` | OpenAPI 3.0 spec + Swagger UI |
 | `flash-extensions/flash-ext-oidc` | OIDC Authorization Code + PKCE flow |
@@ -14,7 +15,6 @@ A high-performance HTTP/1.1 server library for Java 21, built around virtual thr
 | `flash-extensions/flash-ext-view-core` | Minimal shared SSR runtime primitives |
 | `flash-extensions/flash-ext-view-jte` | Opinionated jte SSR extension |
 | `flash-extensions/flash-ext-view-thymeleaf` | Opinionated Thymeleaf SSR extension |
-| `flash-bench` | Demo harness (OIDC + OpenAPI + Jackson) |
 
 ## Requirements
 
@@ -58,7 +58,7 @@ app.post("/echo", (req, res) -> {
 });
 
 app.get("/users/{id}", (req, res) -> {
-    String id = req.pathParam("id");
+    String id = req.param("id");
     return "user:" + id;
 });
 ```
@@ -167,13 +167,64 @@ app.onException((ex, req, res) -> {
 | `tls` | `null` | TLS for the default listener — see [TLS](#tls) |
 | `listeners` | `[]` | Multiple bind targets (port + host + optional TLS) on one app — see [TLS](#tls) |
 | `maxHeaderBufferSize` | `65536` | Max size of the header buffer (bytes) |
+| `wsFrameBufferSize` | `65536` | Per-connection WebSocket read buffer (bytes) |
+| `headerReadTimeoutMs` | `10000` | Once a request's first byte arrives, how long the full header block may take. Bounds slowloris-style attacks — see [`HTTP1-HARDENING.md`](flash/docs/core/HTTP1-HARDENING.md). |
+| `idleKeepAliveTimeoutMs` | `60000` | How long a keep-alive connection may sit idle waiting for its next request. |
+| `bodyReadTimeoutMs` | `30000` | How long reading a request body (handler or automatic drain) may take. |
+| `shutdownDrainTimeoutMs` | `15000` | How long graceful shutdown waits for in-flight requests before force-closing. |
+| `maxConnections` | auto (~heap/10MB) | Maximum concurrent connections across all listeners before new ones are closed immediately at accept time, before any per-connection state (TLS handshake included) is created. Auto-scales from `Runtime.maxMemory()`; set explicitly for a known deployment size, or `0` to disable. |
+| `http2Enabled` | `false` | Whether TLS listeners advertise HTTP/2 through ALPN. |
+| `http2CleartextEnabled` | `false` | Whether plaintext listeners accept HTTP/2 prior knowledge (h2c). Independent from TLS HTTP/2. |
+| `h2HuffmanDynamicValues` | `false` | HPACK-Huffman encode runtime response values. Constants remain pre-encoded; the measured default avoids an extra encode pass. |
+| `h2MaxResetStreamsPerInterval` | `200` | Rapid Reset budget per rolling interval. |
+| `h2MaxStreamsCreatedPerInterval` | `400` | New-stream budget per rolling interval. |
+| `h2AbuseRateIntervalMs` | `10000` | Rolling interval for the two operator-tunable rate limits above. |
+| `h2MaxStreamsPerConnection` | `100000` | Total stream budget; `0` disables it. |
+| `h2MaxBytesPerConnection` | `0` | Optional total wire-byte budget; `0` disables it. |
+| `h2MaxConnectionLifetimeMs` | `0` | Optional connection lifetime; `0` disables it. |
+| `h2StreamIdleTimeoutMs` | `60000` | Inactive open-stream deadline. |
+| `sendDate` | `true` | Add an RFC 9110 `Date` field to responses; disable when an upstream proxy supplies it. |
+
+## Protocols
+
+Routes, middleware, `Request`, `Response`, bodies, trailers, streaming and WebSockets use the same
+API on HTTP/1.1 and HTTP/2. Protocol selection happens once per connection:
+
+- On TLS listeners, enable `http2Enabled`; Flash advertises `h2` and `http/1.1` through ALPN and
+  uses the protocol selected by the client. Existing HTTP/1.1 clients continue to work.
+- On plaintext listeners, enable `http2CleartextEnabled` to accept the HTTP/2 prior-knowledge
+  preface on the same port as HTTP/1.1. Clients that do not send that exact preface are parsed as
+  HTTP/1.1.
+- With both switches left at their default `false`, Flash behaves as an HTTP/1.1 server.
+
+After enabling the appropriate switch, application routes need no protocol-specific code. TLS
+still requires the normal certificate configuration shown below.
+
+Flash deliberately does not implement HTTP/2 server push, RFC 7540 dependency-tree priority
+scheduling, or the obsolete HTTP/1.1 `Upgrade: h2c` transition. Server push has no application API,
+RFC 9113 deprecated the old priority scheme, and cleartext HTTP/2 uses prior knowledge instead.
+See the [HTTP/2 compliance record](flash/docs/http2/COMPLIANCE.md) for exact coverage.
+
+## WebSockets over HTTP/2
+
+The same `ws(path, handler)` route serves WebSockets over HTTP/1.1 and HTTP/2. When HTTP/2 is
+enabled, Flash advertises RFC 8441 extended CONNECT support and carries WebSocket frames inside
+flow-controlled DATA frames. No alternate handler, route, or session API is required:
+
+```java
+app.ws("/live", handler);
+```
+
+HTTP/1.1 clients use the ordinary `101 Switching Protocols` upgrade. HTTP/2 clients use an
+extended CONNECT and receive status `200`; Flash applies the same RFC 6455 framing, masking,
+fragmentation, close, and callback behavior on both transports. Client support for negotiating
+WebSockets over HTTP/2 varies, so clients without RFC 8441 support continue to use HTTP/1.1.
 
 ## TLS
 
-HTTPS and WSS are a transport-layer concern only: once a listener is bound, the accepted
-`Socket` is either plain or an `SSLSocket` indistinguishably from `HttpServer`'s point of view
-onward — the request parser, router, and WebSocket upgrade never branch on it. WSS is therefore
-not a separate feature; it's a WebSocket upgrade running over whatever transport it was handed.
+HTTPS and WSS are a transport-layer concern only. Once a listener is bound, the accepted socket
+is plain or TLS; the selected HTTP connection implementation then performs either the HTTP/1.1
+upgrade or the HTTP/2 extended CONNECT. WSS does not require a separate route or handler API.
 
 ### Quick start
 
@@ -254,23 +305,108 @@ that got the request this far has already completed, never a forced handshake.
 `WebSocketSession` mirrors this exactly (`isSecure()`, `sslSession()`) by delegating to the
 upgrading `Request` — no separate TLS state is tracked for WS.
 
+## Object lifetime
+
+`Request` and `Response` are **pooled per connection**, not allocated per request: one instance is
+created per connection and repositioned (`reset()`) over each new request/response in turn — the
+same idiom Java NIO buffers use, applied to the whole request/response model
+(`flash/docs/core/MESSAGE-MODEL.md` has the full design record). This is what makes a warm h1
+request/response cycle 0 B/op.
+
+**Do not retain a `Request` or `Response` past the handler that received it.** A reference kept in
+a field, a captured closure, a `CompletableFuture` continuation, or a background thread and read
+*after* the handler returns will observe whatever the *next* request on that connection
+repositioned the same instance to — not the request you thought you had:
+
+```java
+// WRONG — captures `req`, reads it after the handler has returned
+app.get("/slow", (req, res) -> {
+    CompletableFuture.runAsync(() -> log(req.header("X-Trace-Id"))); // may log the NEXT request's header
+    return "ok";
+});
+```
+
+Copy out whatever you need before returning or handing work off asynchronously — every accessor
+that returns a `String` (`header`, `param`, `query`, `path`, …) gives you an independent heap copy
+that's safe to keep as long as you like:
+
+```java
+app.get("/slow", (req, res) -> {
+    String traceId = req.header("X-Trace-Id");             // copy now, safe to retain
+    CompletableFuture.runAsync(() -> log(traceId));
+    return "ok";
+});
+```
+
+Run with `-Dflash.env=dev` and a use-after-return access throws `IllegalStateException` immediately
+at the offending call site instead of silently reading the wrong request's data — turn this on in
+tests and local development. It's a no-op in production beyond a single `boolean` field read.
+
+`req.body()`/`RequestBody` follows the same rule — materialise (`.bytes()`) or fully consume
+(`.stream()`) it inside the handler; don't stash the `RequestBody` itself for later.
+
+### Reusable response headers
+
+Use `PreEncodedHeader` for a constant header sent by many responses. It stores the name and value
+once and remains valid on both HTTP versions:
+
+```java
+private static final PreEncodedHeader NO_STORE =
+    new PreEncodedHeader("cache-control", "no-store");
+
+app.get("/health", (req, res) -> res.header(NO_STORE).body("ok"));
+```
+
+`Response.header(byte[])` accepts a complete CRLF-terminated HTTP/1 field line and is therefore
+HTTP/1-only; HPACK needs the name and value as separate fields. Prefer `PreEncodedHeader` for shared
+application and middleware code.
+
+### Trailers and push streaming
+
+Request trailers become available after the body reaches EOF:
+
+```java
+byte[] payload = req.body().bytes();
+String status = req.trailers().first("grpc-status");
+```
+
+For a producer-driven response, `Response.streaming` provides a blocking `ResponseStream`. Its
+bounded buffer and HTTP/2 flow-control windows apply backpressure directly to the producer's
+virtual thread:
+
+```java
+return res.streaming(stream -> {
+    try {
+        stream.write(payload, 0, payload.length);
+        stream.trailer("result", "complete");
+    } catch (IOException failure) {
+        throw new UncheckedIOException(failure);
+    }
+});
+```
+
+The API renders as chunked data and trailers on HTTP/1.1, and DATA plus trailing HEADERS on
+HTTP/2. Flash core supplies these transport primitives; a higher-level gRPC codec belongs in a
+future `flash-ext-grpc` extension.
+
 ## Architecture
 
 ```
-ServerSocket.accept()
-  → RequestParser.parse()       # zero-alloc header parsing, buffer reuse across keep-alive
-  → GlobalRouter.route()        # two-tier: mounted sub-routers (longest prefix) then FastPathRouterImpl
-  → RequestHandler.handle()     # user handler; return value sets body
-  → Request.drain()             # consume unread body for keep-alive
-  → HttpServer writes response  # status line, headers, then fixed or chunked body
-  → loop or close socket        # based on Connection header
+TransportFactory.create()           # binds every listener, wires the connection runner
+  → AcceptLoop                      # one per listener × accept thread; hands sockets off
+  → ConnectionRunner.accept()       # per-connection setup: TLS handshake, protocol negotiation
+      → ProtocolNegotiator          # ALPN / h2c-preface — decides the protocol once
+      ├─ Http1Connection.run()      # request parser, router, handler, h1 response writer
+      └─ Http2Connection.run()      # frame demux, HPACK, stream dispatch, flow control
+             → RequestHandler.handle()  # the same protocol-neutral request/response API
 ```
 
-- **Virtual threads** — each accepted socket runs on a virtual thread (`Executors.newVirtualThreadPerTaskExecutor()`). Java 21 required.
+- **Virtual threads** — each accepted socket runs on a virtual thread (`Executors.newVirtualThreadPerTaskExecutor()`, owned by `TransportFactory`). Java 21 required.
 - **Zero-allocation router** — `FastPathRouterImpl` uses `fpr-core`, a byte-level FSM that matches on `METHOD + path` bytes with no per-request allocation.
 - **Keep-alive** — `RequestParser` reuses its header buffer across requests on the same connection.
 - **Chunked transfer** — both chunked request bodies (decoded via `ChunkedInputStream`) and chunked response bodies are supported.
 - **TLS is transport-only** — see [TLS](#tls). Listeners bind either a plain `ServerSocket` or an `SSLServerSocket`; nothing downstream of `accept()` branches on which.
+- **`ConnectionProtocol` seam** — HTTP/1.1 and HTTP/2 are peers behind this interface, selected once per connection by `ProtocolNegotiator`; routing and application models are shared.
 
 ## Build & test
 
@@ -283,7 +419,4 @@ mvn test
 
 # Run a single test class
 mvn test -pl flash -Dtest=RequestParserTest
-
-# Run the benchmark demo server
-java -jar flash-bench/target/flash-bench-1.0-SNAPSHOT.jar
 ```

flash/docs/core/BYTES.md

@@ -0,0 +1,177 @@
+# The byte layer
+
+Audience: contributors. This is the design record for `dev.relism.flash.bytes` — the
+protocol-neutral byte primitives both HTTP/1.1 and HTTP/2 build on — and for the Phase 4
+allocation/scanning fixes (`EX-04`, `EX-05`, `EX-09`, `EX-19`, `EX-25`, `EX-26`, `EX-33`, plus
+`EX-06`'s router half) that consume them.
+
+## Why this exists
+
+Before Phase 4, byte-scanning and case-insensitive comparison logic was duplicated, slightly
+differently, in `RequestParser`, `HeaderMap`, and `Http1KeepAlive`; `fpr-core`'s word-at-a-time
+router-matching fast path (`ByteCompare`) was wired up but never actually enabled anywhere
+(`EX-04` — every `ByteView` implementation returned `supportsLong() == false`); and four call
+sites allocated a fresh view, array, or `String` per call on paths a realistic middleware chain
+hits 6–10 times per request. `dev.relism.flash.bytes` is the single home these fixes converge on,
+so no later phase (HPACK, the frame layer) has to invent its own scanning primitives.
+
+## Package layout
+
+```
+dev.relism.flash.bytes
+├── ByteScan                 static scanning/comparison/hashing utilities, scalar + SWAR
+├── ArrayBackedByteView       capability interface: a ByteView backed by one contiguous byte[]
+├── SegmentedByteView         the deliberate non-array-backed case (K discontiguous segments)
+├── PooledSlice                reusable ArrayBackedByteView, the EX-05 fix
+├── SlicePool                   a small fixed-size ring of PooledSlice
+├── ByteWriter                 index-based writer into a growable byte[] scratch buffer
+└── Pairs                       the (hi<<32)|lo allocation-free pair-return idiom, named
+```
+
+## The `ByteView` capability hierarchy
+
+```
+ByteView                              (fpr-core)
+├── ArrayBackedByteView                capability: array() + offset()
+│   ├── FastPathViews.RequestByteView  RequestParser's request-line/header slices
+│   ├── FastPathViews.SocketByteView   a bare byte[] (e.g. a WebSocket payload)
+│   ├── FastPathViews.StringByteView   a String's UTF-8 bytes
+│   └── PooledSlice                    the EX-05 reusable, pool-issued slice
+└── (bare ByteView, not array-backed)
+    ├── SegmentedByteView               K discontiguous segments (general-purpose; HPACK stays contiguous)
+    └── FastPathViews.MethodPathByteView  method bytes + another ByteView, composed
+```
+
+Code holding a bare `ByteView` and wanting the fast path when the concrete instance happens to
+be array-backed does `instanceof ArrayBackedByteView` and falls back to the byte-at-a-time path
+otherwise — see `ArrayBackedByteView`'s own Javadoc. This is used throughout Phase 4's fixes:
+`Request.path()`, `PathParams.get()` (`EX-25`), and `QueryParams.decode`'s clean-value fast path
+(`EX-26`) all take this shape.
+
+## `EX-04`: the `supportsLong()`/`longAt()` contract
+
+`fpr-core`'s `ByteCompare` (decompiled from `fpr-core-1.1.1`, since no source jar is published)
+reads its comparison word via
+`MethodHandles.byteArrayViewVarHandle(long[].class, ByteOrder.LITTLE_ENDIAN)` and only takes the
+word-at-a-time branch when the caller passed `useLong = true`, comparing the result bit-for-bit
+against whatever `ByteView#longAt` returns. The contract this imposes on any `longAt`
+implementation:
+
+- Return the same value `LONG_VIEW.get(array, pos)` would, for the identical 8 bytes — meaning
+  **little-endian**, fixed, regardless of the host's native byte order (unlike `ByteScan`'s own
+  SWAR internals, which use `ByteOrder.nativeOrder()` for speed — see below for why that's a
+  different, safe choice in a different context).
+- The caller (`ByteCompare`) never calls `longAt(i)` without first establishing `i + 8 <=
+  length()` — so `longAt` implementations do not re-check this themselves (a defensive check
+  would be dead code on every real call path).
+
+`FastPathViews.RequestByteView`/`SocketByteView`/`StringByteView` implement this;
+`MethodPathByteView` (composite, no single backing array) and `SegmentedByteView` (genuinely
+discontiguous) both stay at the inherited `false` default — a word-at-a-time read is not merely
+unimplemented for these, it is structurally unsound (a read could straddle two sources).
+
+**Verified against `fpr-core` directly** (`FastPathViewsLongAtTest`), not merely by reading
+bytecode: `ByteCompare.equals`/`indexOf` called with `useLong=true` and `useLong=false` are
+asserted to agree on identical content, on content diverging at every position across an
+8+-byte range (word-interior, word-boundary, and scalar-tail cases), and end-to-end through a
+real compiled `fpr-core` router with literal route segments ≥ 8 bytes — including a near-miss
+route differing only in its last byte, to catch exactly the kind of bounds/endianness bug that
+would otherwise silently mis-route a request (the failure mode `EX-04`'s registry entry calls out
+by name as the worst possible one here).
+
+## `ByteScan`'s SWAR technique
+
+Both `ByteScan.indexOf` (single byte) and `ByteScan.indexOfCrLfCrLf` (the `\r\n\r\n` header
+terminator, `EX-33`) use the classic "does this word contain byte `b`" bit trick: XOR the 8-byte
+word against `b` broadcast into every lane, then test for any zero lane with
+`(v - 0x0101...01) & ~v & 0x8080...80`. `indexOfCrLfCrLf` uses this as a pre-filter to find a
+candidate `CR` byte 8 at a time, then a cheap scalar 3-byte check verifies the full 4-byte match
+at each candidate — so a scan touches every byte once per 8-byte stride in the common
+no-CR-yet case, rather than once per byte.
+
+This reads the word via `ByteOrder.nativeOrder()`, not a fixed order — safe here (unlike
+`EX-04`'s `longAt`) because nothing compares this word against an independently-decoded one;
+byte-equality detection itself (finding *that* a matching lane exists) is indifferent to lane
+order, and position extraction (`laneIndexOf`) branches on the actual native order once, at
+class-init time, to convert a matching bit back into the correct array index either way.
+
+Every SWAR method has a scalar counterpart (`indexOfScalar`, `indexOfCrLfCrLfScalar`) used as
+the correctness oracle: `ByteScanTest` property-tests SWAR against scalar at every length 0–256
+and every match position (including unaligned starts and matches at the very last valid byte),
+and `ByteScanFuzzTest` throws 20 000 fully-random trials at each, per the plan's task 1. All
+green — see the class's own Javadoc for the full technique writeup.
+
+## `Http1HeaderMap`'s index
+
+Originally, every `Http1HeaderMap` lookup (`first`, `all`, `view`, `valueEqualsIgnoreCase`)
+rescanned the entire header section from scratch — O(n·m) for a realistic middleware chain
+performing 6–10 lookups per request. `RequestParser` now populates the index while it validates
+each header line; direct `Http1HeaderMap.reset()` callers scan the section exactly once. It records
+per-header `(nameOffset, nameLength, valueOffset, valueLength)` and a case-insensitive
+32-bit FNV-1a hash of the name (`ByteScan.hashNameIgnoreCaseAscii`) into `int[]` arrays grown
+(never shrunk) to the connection's high-water mark, capped by `Http1Limits.MAX_HEADER_COUNT`
+(asserted, not silently truncated — the parser rejects a request that would exceed it).
+Every lookup then compares the caller's own hash (`ByteScan.hashNameIgnoreCaseAscii(String)`,
+computed once) against the index's hashes before ever falling back to a full case-insensitive
+name comparison. Production therefore performs one combined validation/index pass rather than
+one parse-time pass plus one rescan per lookup. `forEach` uses the same index rather than keeping
+an independent scanner.
+
+## `EX-05`: pooled slices
+
+`Http1HeaderMap.view`, `QueryParams.view`, and `PathParams.view` used to allocate a fresh anonymous
+`ByteView` (plus its capturing instance) on every call. Each now draws from a small
+(`VIEW_POOL_SIZE = 4`) `SlicePool` of reusable `PooledSlice` instances instead. The lifetime
+contract, restated on each method: **a returned view stays valid until either the request ends,
+or the same `view()` method is called `VIEW_POOL_SIZE` more times on the same instance —
+whichever comes first** — at which point the ring silently repositions the same object over
+different bytes. This is a real, demonstrated hazard, not a hypothetical one:
+`SlicePoolTest#wraparoundAliasesThePreviouslyReturnedSlice` and the analogous tests in
+`Http1HeaderMapIndexTest`, `QueryParamsFastPathTest`, and `PathParamsTest` all show a 5th call
+returning the exact same object instance the 1st call did, now aliased to different content.
+
+`QueryParams` and `PathParams`'s pools are created **lazily**, on the first actual `view()` call
+— not eagerly in the constructor — because both classes are otherwise-cheap objects created per
+request (or, for `PathParams`'s `FastPathRouterImpl`-owned reusable instance, once per
+connection) regardless of whether `view()` is ever invoked; an eager pool would add
+`VIEW_POOL_SIZE` allocations to every such object whether or not it needed them; `Http1HeaderMap`'s
+pool, by contrast, is unconditionally useful (every request's map handles headers) and is
+constructed eagerly for simplicity.
+
+**Two documented, deliberately-kept exceptions to "no `new ByteView()` remains"**: `QueryParams.view`
+and `PathParams.view` each retain a fallback anonymous `ByteView` for the case where their
+backing source is *not* `ArrayBackedByteView` — structurally unreachable on the real request path
+today (`RequestParser` only ever constructs array-backed views), kept because both constructors
+are `public` and could in principle be called with an arbitrary `ByteView`. A silent, correct,
+allocating fallback was judged preferable to either crashing on a technically-valid input or
+deleting a case that only test code could exercise. `Http1HeaderMap.view` has no such fallback
+— it is always buffer-backed by construction.
+
+## `EX-19`/`EX-06` (router half): the `FastPathRouterImpl` scratch
+
+`FastPathRouterImpl.RouteScratch` (created once per connection via `AbstractRouter#newScratch`,
+replacing the `ThreadLocal<MatchResult>`/`ThreadLocal<MethodPathByteView>` pair, as an opaque
+caller-owned object rather than an extension of `ConnectionScratch`) also owns the reusable
+path-param arrays and a single long-lived
+`PathParams` instance, grown (via `ensureParamCapacity`, doubling) to the largest param count any
+route on that connection has ever matched, and repositioned (`PathParams#reset`) rather than
+reallocated on every match. `PathParams` gained a second, count-explicit constructor and a public
+`reset(ByteView, int)` specifically for this: the reusable arrays can be larger than a given
+request's actual param count, so `count` must be tracked independently of `names.length`.
+
+## `EX-25`/`EX-26`: single-allocation `String` construction
+
+`Request.path()`, `PathParams.get()`, and (for the common "no `%`/`+` in the value" case)
+`QueryParams.decode` now build their result `String` directly from the backing array via
+`new String(array, offset, length, UTF_8)` when the source is `ArrayBackedByteView`, instead of a
+byte-at-a-time copy into a scratch `byte[]` followed by a second allocation for the `String`
+itself. `QueryParams.decode` scans the value once for `%`/`+` first; only a value that actually
+needs percent-decoding pays for the scratch-buffer path — verified to produce byte-identical
+output to the always-decode path it bypasses, across clean values, `+`-only, `%XX`-only, invalid
+escapes, and mixed queries (`QueryParamsFastPathTest`).
+
+## Performance measurement
+
+`EX-04` (the router's word-at-a-time path) and `EX-33` (the SWAR header-end scan) both carried an
+explicit "measure, and keep only if it doesn't cost" requirement. Both were measured together
+with the phase's overall zero-allocation contract in one JMH pass, and both were kept.

flash/docs/core/HTTP1-HARDENING.md

@@ -0,0 +1,92 @@
+# HTTP/1.1 hardening
+
+Audience: operators. This is the document to read when a `400`/`413`/`414`/`431`/`501` shows up
+in the logs and it isn't obvious why. Every rejection rule Flash's HTTP/1.1 parser enforces is
+listed here with its RFC citation and the status it produces. Contributor-level detail (why each
+check is implemented the way it is, the exact code paths) lives in the Javadoc of
+`RequestParser`, `ChunkedInputStream`, and `dev.relism.flash.exceptions.MalformedRequestException`.
+
+Every rejection in this document has one thing in common: **the connection is always closed
+afterwards, never kept alive.** A rejected request is exactly the situation a smuggling attack
+needs a reusable connection for, so none of these rejections offer one — see
+`MalformedRequestException`'s Javadoc.
+
+## Request-smuggling defenses (RFC 9112 §6.1)
+
+| Rule | Status | Detail |
+|---|---|---|
+| `Content-Length` and `Transfer-Encoding` both present | `400` | The canonical CL.TE/TE.CL smuggling vector. Rejected regardless of which header appears first. |
+| Multiple `Content-Length` lines with **differing** values | `400` | Identical repeated values are tolerated (RFC 9110 §8.6 permits treating them as one). |
+| `Transfer-Encoding` whose **final** coding is not `chunked` | `501` | Flash implements only `chunked`; anything else (`gzip` alone, or `chunked, gzip` — chunked must be *last*) is unsupported. |
+
+## Strict `Content-Length` parsing (RFC 9110 §8.6)
+
+| Input | Status |
+|---|---|
+| Empty value | `400` |
+| Any non-digit byte (including a leading `+` or `-`) | `400` |
+| More than 19 digits | `400` |
+| Value overflows `Long.MAX_VALUE` | `400` |
+| Value exceeds `Http1Limits.MAX_CONTENT_LENGTH` (4 GiB by default) | `413` |
+
+The previous parser silently skipped non-digit characters (`"5abc"` parsed as `5`; `"-1"` parsed
+as `1`) instead of rejecting them — this is the fix.
+
+## Header and request-line limits (`Http1Limits`)
+
+| Limit | Default | Status when exceeded |
+|---|---|---|
+| `MAX_HEADER_COUNT` | 100 | `431 Request Header Fields Too Large` |
+| `MAX_HEADER_NAME_LENGTH` | 256 B | `431` |
+| `MAX_HEADER_VALUE_LENGTH` | 8192 B | `431` |
+| `MAX_REQUEST_LINE_LENGTH` | 8192 B | `431` |
+| Header block exceeds `maxHeaderBufferSize` (or the connection ends before it completes) | configurable, default 64 KiB | `431` |
+
+## Line-terminator and header-syntax correctness (RFC 9112 §5)
+
+| Rule | Status |
+|---|---|
+| A `\r` not immediately followed by `\n` (bare CR) | `400` — a known desynchronization/smuggling surface |
+| A header line beginning with whitespace (obsolete line folding, RFC 9112 §5.2) | `400` |
+| A header name containing a byte outside RFC 9110 §5.6.2's `tchar` set | `400` |
+| A header line with no `:` | `400` |
+
+## Chunked transfer safety (RFC 9112 §7.1, `Http1Limits`)
+
+| Limit | Default | Status when exceeded |
+|---|---|---|
+| `MAX_CHUNK_SIZE` | 16 MiB | `413` |
+| Chunk-size line longer than 16 hex digits | — | `400` |
+| `MAX_CHUNK_EXT_LENGTH` (the optional `;name=value` after a chunk size) | 256 B | `400` |
+| `MAX_CHUNKS_PER_BODY` | 100 000 | `413` |
+| `MAX_TRAILER_COUNT` | 50 | `431` |
+| A chunk's data not followed by `\r\n`, or a malformed chunk-size/trailer terminator | — | `400` |
+
+Trailers are consumed within the bounds above and exposed through `Request.trailers()` on both
+HTTP/1.1 and HTTP/2.
+
+## Timeouts (`FlashConfiguration`)
+
+| Setting | Default | Covers |
+|---|---|---|
+| `idleKeepAliveTimeoutMs` | 60 000 | How long a keep-alive connection may sit idle waiting for its next request. |
+| `headerReadTimeoutMs` | 10 000 | Once the first byte of a request arrives, how long the full header block may take. |
+| `bodyReadTimeoutMs` | 30 000 | How long reading the body (by the handler, or the automatic post-response drain) may take. |
+| `shutdownDrainTimeoutMs` | 15 000 | How long graceful shutdown waits for in-flight requests before force-closing. |
+
+These are enforced by an **absolute deadline**, not merely `Socket.setSoTimeout`. A per-read
+socket timeout alone never trips against a peer that sends one byte just often enough to keep
+each individual read alive (the classic slowloris shape) — see
+`dev.relism.flash.transport.BufferedByteSource`'s Javadoc for how the absolute deadline is
+implemented on top of the JDK's per-read-only timeout API.
+
+## TLS (RFC 9113 §9.2.2, applies once a listener offers `h2` over ALPN)
+
+- The TLS handshake is forced explicitly (not left to the JDK's lazy on-first-read trigger)
+  before any protocol decision is made, and is bounded by `headerReadTimeoutMs`.
+- When a listener's `TlsConfig.applicationProtocols` includes `"h2"`, the enabled TLS 1.2 cipher
+  suite list is filtered against the RFC 9113 Appendix A blocklist
+  (`TlsConfig.TLS12_H2_BLOCKED_CIPHERS`, ~280 entries). TLS 1.3 is never affected — none of its
+  cipher suites are on that list.
+- `FlashConfiguration.http2Enabled` advertises `h2` on TLS listeners. The independent
+  `http2CleartextEnabled` switch accepts the h2c prior-knowledge preface on plaintext listeners.

flash/docs/core/MESSAGE-MODEL.md

@@ -0,0 +1,191 @@
+# The message model
+
+Audience: contributors. This is the design record for `dev.relism.flash.models`'s shared
+request/response model: what is pooled, what that pooling means for callers, and how HTTP/1.1 and
+HTTP/2 retain the same public contract.
+
+## Why this exists
+
+Through Phase 5, `Request`, `RequestLine`, `RequestBody`, and `Response` were all allocated fresh
+per request — `DEC-20` measured this at 120.008 B/op for parse+route alone, and traced 100% of it
+to these four objects. Phase 6 pools all of them, following the same "one instance per connection,
+repositioned via `reset()`, never reallocated" idiom `Http1HeaderMap` and `RequestLine` already
+established in earlier phases. This document is the single place that idiom's contract — and the
+hazards of misusing it — is written down for the whole model, instead of being re-derived from
+each class's own Javadoc.
+
+## What is pooled, and by whom
+
+```
+RequestParser (one per connection)
+├── Http1HeaderMap  headerMap    — reset() per request
+├── RequestLine     requestLine  — reset() per request
+├── Request         request     — reset() per request (via Request.forParsed)
+├── RequestBody     requestBody — reset() per request
+├── RequestByteView pathView     — reset() per request                    (EX-42)
+├── RequestByteView queryView    — reset() per request, only when present (EX-42)
+└── RequestByteView protocolView — reset() per request                    (EX-42)
+
+Http1Connection (one per connection)
+└── Response pooledResponse — reset() per request (unless a handler returns its own Response)
+
+FastPathRouterImpl.RouteScratch (one per connection, via AbstractRouter#newScratch)
+└── PathParams pathParams — reset() per matched request (see BYTES.md, EX-19)
+```
+
+Every one of these follows the same three rules:
+
+1. **One instance per connection**, created once (in `RequestParser`'s or `Http1Connection`'s
+   constructor, or in `newScratch()`), never re-allocated for the connection's lifetime except a
+   backing array growing to a new high-water mark (e.g. `RequestParser.buffer` doubling, or
+   `RouteScratch.ensureParamCapacity`).
+2. **`reset(...)` repositions, it does not allocate** — the method that transitions the instance
+   from "describes request N" to "describes request N+1".
+3. **Do not retain past the handler.** A reference captured in a closure, a `CompletableFuture`
+   continuation, or a background thread and read after the handler returns will observe whatever
+   the *next* request repositioned the instance to — silently, unless the dev-mode guard below
+   catches it.
+
+## The dev-mode use-after-recycle guard (`Request`, `Response`)
+
+`Request` and `Response` — the two objects most likely to be captured by user code — additionally
+track an `active` flag, set `true` by `reset()` and `false` by `recycle()` (called by
+`Http1Connection` once the handler and `drain()` have finished). Every public accessor calls
+`checkActive()` first:
+
+```java
+private void checkActive() {
+    if (poisoningEnabled && !active) {
+        throw new IllegalStateException("... do not retain a Request past the handler ...");
+    }
+}
+```
+
+`poisoningEnabled` defaults to `Flash.DEV` (`-Dflash.env=dev`), so this is a zero-cost `static
+final`-guarded branch in production and a loud, precise `IllegalStateException` — thrown at the
+exact misusing call site — in development. Since `Flash.DEV` is itself `static final` (fixed at
+JVM startup) and therefore not something a single test can toggle, both classes expose a
+package-private `setPoisoningEnabledForTesting(boolean)` hook purely so
+`RequestRecycleGuardTest`/`ResponseRecycleGuardTest` can exercise the dev-mode branch without a
+fragile reflective override of a `static final` field — production code never touches it.
+
+`RequestBody`, `RequestLine`, `Http1HeaderMap`, and `PathParams` do **not** carry this guard: they
+are reached only through `Request`/`Response` (or, for `PathParams`, through `Request.param`),
+so `Request`/`Response`'s own guard already catches a stale read before it would reach these.
+
+## `RequestBody`: two read modes, one reused bounded stream
+
+`RequestBody.stream()` and `.bytes()` are mutually exclusive per request (calling both is
+undefined). `EX-23`/`EX-24` (Phase 6) replaced two allocation sources in the streaming path:
+
+- `stream()` used to build a fresh `SequenceInputStream` + `ByteArrayInputStream` + anonymous
+  bounded `InputStream` on every call. It now repositions one persistent
+  `BoundedBufferedInputStream` (a private inner class) via `reset(preBuf, preBufOff, preBufLen,
+  socketRemaining)` — the same object is returned every time, just pointed at different bytes.
+- `drain()`'s chunked-body path used to call `InputStream.transferTo`, whose default
+  implementation allocates a fresh 8 KiB `byte[]` on every call. It now drains through a lazily
+  created (only if a chunked body is ever actually drained), persistent `drainBuffer`.
+
+`RequestBody.of(byte[])`/`.empty()` remain as freestanding, unpooled factories for test/manual
+construction (mirroring `Request`'s own manual constructor) — production's only pooled instance is
+the one `RequestParser` owns.
+
+## `Response`: byte-level headers, one write, `ResponseSerializer` as the source of truth
+
+Before Phase 6, `Response.header(String, String)` stored headers as `List<byte[]>` — one `String`
+concatenation and one `byte[]` allocation per call. `Response` now stores structured headers in a
+`ByteWriter`-backed name/value region plus parallel `int[]` quads (`nameOff, nameLen, valOff,
+valLen`), written via `ByteWriter.writeAscii` — zero-allocation on a warm connection. A second,
+separate store (`List<byte[]>`) still holds the legacy `header(byte[])` raw-line entries; a tagged
+sequence (`headerTags`/`headerRefs`) interleaves the two stores back into declaration order when
+serialized, so mixing `header(String,String)` and `header(byte[])` calls on the same response still
+produces headers in the order they were added.
+
+`PreEncodedHeader` precomputes a header's name and value ASCII bytes once (for example, a constant
+response header set at boot). Preserving the boundary lets HTTP/1.1 render a field line and HTTP/2
+encode the same pair through HPACK without a second public header type.
+
+`ResponseSerializer.forEachField(Response, FieldConsumer)` is the **one source of truth for what
+headers a response has** — it enumerates `Content-Type` (if set) plus every structured custom
+header, in order, and is the only place that knowledge lives. `Http1ResponseWriter` renders that
+sequence as `Name: Value\r\n` lines; `Http2ResponseWriter` renders the same sequence as HPACK.
+Deliberately excluded: `Content-Length`/`Connection`/`Date` (connection framing, not response
+object properties — and HTTP/2 has no `Connection` header at all, RFC 9113 §8.2.2) and raw
+`header(byte[])` entries (no recoverable name/value structure to hand the h2 encoder).
+
+`Http1ResponseWriter` (`EX-27`) serializes the entire response head — status line, `Content-Type`,
+`Date`, every custom header, `Content-Length`/`Connection` — into
+`ConnectionScratch.responseHead` (a reused `ByteWriter`) and issues **one** `OutputStream.write`
+call for the head plus any body at or below `Http1Limits.INLINE_BODY_THRESHOLD` (8 KiB), instead
+of roughly ten small writes. A larger body is written in a second `write` call right after — folding
+it into the head buffer first would cost an extra full-body `memcpy` the syscall reduction does not
+pay for. Streaming/chunked bodies write the head, then relay their own bytes as they arrive, by
+definition too large or unbounded to fold into one buffer up front.
+
+`Response.header(...)` (any overload) is bounded by `Http1Limits.MAX_RESPONSE_HEADER_BYTES`/
+`MAX_RESPONSE_HEADER_COUNT` (`EX-43`) — unlike every other `Http1Limits` constant, this guards
+against a bug in the *caller* (a handler looping over an unbounded collection while building
+headers) rather than a hostile peer: since `Response` is now pooled per connection, an unbounded
+`headerRegion` would otherwise grow for the rest of the connection's lifetime, never shrinking
+back down between requests. Both checks throw `IllegalStateException`, not
+`MalformedRequestException` — this is an application-code misuse, not a wire-input rejection.
+
+## `HeaderView` / `Http1HeaderMap` (`DEC-22`)
+
+`HeaderMap` split into `HeaderView` (the protocol-neutral read contract: `first`, `all`, `view`,
+`valueEqualsIgnoreCase`, `contains`, `count`, `forEach`) and `Http1HeaderMap` (the existing
+byte-buffer-backed implementation, kept in `dev.relism.flash.models` rather than moved to
+`dev.relism.flash.http1`: `RequestParser` (root package) owns and constructs it, and
+`http1`→root already exists via `Http1Connection`, so moving it to `http1` would create a
+`models`↔`http1` package cycle). `RequestLine.headers` is typed as the
+interface; `Http2HeaderMap` is the HPACK-backed second implementation without requiring a
+`Request` or `RequestLine` API split.
+
+## `ByteTemplate` (`EX-28`)
+
+Off the h1 request/response hot path (used only by `ErrorPages`, on 404/500), but in scope because
+it was a clean instance of the "precompute at boot" category the phase's own text calls out.
+`render(String...)` used a nested loop — for every key-value pair, scan every slot — to find
+matching placeholders, and a repeated placeholder name (`{{var}} == {{var}}`) meant a naive
+name→single-index map would be wrong. Fixed by mapping each slot name to the (usually
+one-element) array of every slot index using that name, built once at construction. A new
+`renderInto(byte[], int, String...)` overload writes into a caller-supplied buffer and returns the
+length written, for future callers with a reusable scratch buffer available; `render(String...)`
+keeps its allocating signature for compatibility.
+
+## `Multipart` (`EX-29`, and `EX-38`–`EX-41`)
+
+Audited per the plan's mandatory rules for any file over 300 lines. Findings and fixes: an eagerly
+buffered part body (text fields, and — during a full `parts()`/`parts(String)` scan — file bodies
+too) had no size bound (`EX-38`, fixed with a bounded read capped by
+`Http1Limits.MAX_MULTIPART_BUFFERED_PART_SIZE`); the part count was unbounded (`EX-39`, capped by
+`Http1Limits.MAX_MULTIPART_PARTS`); per-part header parsing had neither a header-count nor a
+line-length bound (`EX-40`, capped by `Http1Limits.MAX_MULTIPART_PART_HEADER_COUNT`/
+`MAX_MULTIPART_HEADER_LINE_LENGTH`); the multipart boundary's length was checked and found to
+already be bounded transitively, via `Http1Limits.MAX_HEADER_VALUE_LENGTH` on the `Content-Type`
+header it comes from (`EX-41`, a non-finding, recorded so "checked, found fine" isn't mistaken for
+"wasn't checked"). None of these bounds apply to `Part.materialize()` on a streaming file part
+returned by `Multipart.file()` — that call is documented as an explicit, opt-in heap allocation the
+caller chooses to pay for, the same way `RequestBody.bytes()` is.
+
+## `EX-42`: the last per-request allocation, found by re-measuring
+
+Pooling `Request`/`RequestBody`/`RequestLine`/`Response` dropped `RequestPipelineBenchmark`'s
+`parseAndRoute` from 120.008 B/op to 48.008 B/op — real progress, but not the 0 B/op the phase's
+own DoD text requires. Reading `RequestParser.parse` turned up three `new
+FastPathViews.RequestByteView(...)` allocations (path, query when present, protocol) on every
+call — pre-existing since at least Phase 4, invisible until the larger `Request`/`RequestBody`/
+`RequestLine` cost sitting on top of them was removed. Fixed the same way as everything else in
+this document: `RequestByteView` gained a `reset(byte[], int, int)`; `RequestParser` now owns one
+pooled instance per role. `parseAndRoute` measures 0.008 B/op after the fix — JMH's noise floor,
+effectively 0.
+
+## The zero-alloc contract, closed
+
+> A complete h1 request/response cycle on a warm connection — parse, route with path params, read
+> three headers, set two response headers, write a 200 with a `byte[]` body — must be 0 B/op.
+
+`RequestPipelineBenchmark.parseAndRoute` (parse + route with a parametric match, no header/param
+access) measures 0 B/op. `parseRouteAndExtractThreeFields` (the same, plus one path param and two
+header reads) measures 184.009 B/op — entirely the `String` allocations the contract's own text
+exempts ("except for the user-facing `String`s the handler explicitly asks for").

flash/docs/core/README.md

@@ -0,0 +1,10 @@
+# Flash core
+
+The parts of Flash shared by every protocol it speaks — HTTP/1.1 and HTTP/2 alike. Protocol-specific
+internals (frames, HPACK, stream state) live in [`../http2/`](../http2/README.md).
+
+- [HTTP/1.1 hardening](HTTP1-HARDENING.md) — message-boundary rules, timeouts and negotiation.
+- [Transport](TRANSPORT.md) — listeners, connection ownership, TLS and virtual threads.
+- [Message model](MESSAGE-MODEL.md) — shared request/response objects and their lifetime contract.
+- [Trailers and streaming](TRAILERS-AND-STREAMING.md) — the public cross-protocol APIs.
+- [Byte primitives](BYTES.md) — reusable views, scanning and bounded slice lifetimes.

flash/docs/core/TRAILERS-AND-STREAMING.md

@@ -0,0 +1,36 @@
+# Trailers and streaming
+
+Flash exposes the same request and response model on HTTP/1.1 and HTTP/2. Request trailers are
+available through `Request.trailers()` after the body has reached EOF. Calling it earlier throws
+`IllegalStateException`; this prevents handlers from observing an incomplete trailer section.
+HTTP/1.1 reads trailers from the final chunk, while HTTP/2 decodes the trailing HEADERS block in
+the connection's existing HPACK context.
+
+Response trailers are added with `Response.trailer(name, value)` or a `PreEncodedHeader`. HTTP/1.1
+uses chunked framing and writes the fields after the zero chunk. HTTP/2 writes a trailing HEADERS
+block with `END_STREAM`; the final DATA frame deliberately does not carry `END_STREAM`.
+
+`Response.streaming(producer)` is the push alternative to `stream(InputStream, length)` and
+`chunked(InputStream)`. Its `ResponseStream` is a bounded blocking bridge. A producer runs on a
+virtual thread and blocks when the protocol writer or the HTTP/2 flow-control windows cannot make
+progress. This keeps backpressure explicit without callbacks or reactive types:
+
+```java
+return response.type("application/grpc").streaming(stream -> {
+  try {
+    for (byte[] message : messages) stream.write(message, 0, message.length);
+    stream.trailer("grpc-status", "0");
+  } catch (IOException failure) {
+    throw new UncheckedIOException(failure);
+  }
+});
+```
+
+The transport supports the primitives required by gRPC, but the core does not provide protobuf
+codecs, generated stubs, service descriptors, or a gRPC service API. Those belong in a future
+`flash-ext-grpc` module. `GrpcInteropTest` verifies the boundary with the external `grpcurl` client
+and a hand-written wire-format handler.
+
+CONNECT requests follow RFC 9113 request pseudo-header rules: `:authority` is required and
+`:scheme`/`:path` are forbidden. Their DATA remains subject to the ordinary request limits,
+timeouts and two-level flow control.

flash/docs/core/TRANSPORT.md

@@ -0,0 +1,148 @@
+# Transport architecture
+
+Audience: contributors. This document describes the shared listener and connection layer behind
+the HTTP/1.1 and HTTP/2 implementations.
+
+## Why this exists
+
+The original `HttpServer` (563 lines) did bind, accept, virtual-thread dispatch, WebSocket
+upgrade detection, WebSocket handshake, the WebSocket session loop, keep-alive detection, HTTP
+response serialization, chunked encoding, hex encoding, and decimal encoding — eleven reasons to
+change in one class (R6). It also held three `ThreadLocal`s that meant "one per connection" under
+virtual threads, not "one per core" (`EX-06`), and used `synchronized` around blocking socket
+writes in `WebSocketSession`, which pins a virtual thread's carrier on Java 21 (`EX-01`).
+
+The current design replaces it with named, single-responsibility components and a
+`ConnectionProtocol` seam implemented by both wire protocols.
+
+## Package layout
+
+```
+dev.relism.flash.transport
+├── TransportFactory        composes everything below; ServerHandle.create()'s implementation (EX-34)
+├── ListenerBinder           FlashConfiguration.Listener -> bound ServerSocket
+├── BoundListener            record: the bound socket + whether it is TLS
+├── TransportTuning          accept-thread count / backlog / socket buffer size constants
+├── AcceptLoop                one listener's accept loop body
+├── ConnectionRunner          per-connection setup/teardown: TLS handshake, protocol negotiation,
+│                             dispatch to a ConnectionProtocol, guaranteed cleanup
+├── ConnectionProtocol        the h1/h2 seam: void run(ConnectionContext)
+├── ConnectionContext         everything a ConnectionProtocol needs, bundled (record)
+├── ConnectionScratch         per-connection reusable buffers (EX-06's fix)
+├── ScratchPool               a bounded cache of ConnectionScratch instances
+├── ServerLifecycle           implements ServerHandle: start/startAndBlock/stop, graceful shutdown (EX-32)
+├── BufferedByteSource        the buffered, deadline-aware, peekable inbound-byte source (Phase 1, EX-10)
+└── ProtocolNegotiator/NegotiatedProtocol   ALPN + h2c preface detection (Phase 1)
+
+dev.relism.flash.http1
+├── Http1Connection            implements ConnectionProtocol: the h1 keep-alive request loop
+├── Http1ResponseWriter        serializes a Response as an HTTP/1.1 message
+└── Http1KeepAlive             keep-alive decision + the shared Connection-header token scanner (EX-13)
+
+dev.relism.flash.websocket (existing package, extended)
+├── WebSocketUpgrade           upgrade detection + handshake response
+├── WebSocketLoop               the session read/dispatch loop
+├── WebSocketSession            per-connection WS I/O (frame codec + send API), EX-01/EX-11/EX-12
+└── WebSocketProtocolException  RFC 6455 violation, carries the correct close code
+```
+
+## The connection lifecycle
+
+```
+TransportFactory.create(configuration, router, wsRouter)
+    binds every configured listener (ListenerBinder)
+    builds one ConnectionRunner (shared virtual-thread executor, ScratchPool, both protocols)
+    returns a ServerLifecycle (implements ServerHandle)
+
+ServerLifecycle.start()
+    for each listener, spawns TransportTuning.ACCEPT_THREADS platform threads
+    each runs AcceptLoop.run(listener, runner, this::isStopped)
+
+AcceptLoop.run(...)
+    loop: listener.socket().accept() -> runner.accept(socket, stopped)
+
+ConnectionRunner.accept(socket, stopped)
+    submits to the virtual-thread executor -> handle(socket, stopped)
+
+ConnectionRunner.handle(socket, stopped)
+    activeSockets.add(socket); scratch = scratchPool.acquire()
+    try:
+        configure TCP_NODELAY / send buffer size
+        if SSLSocket: force startHandshake() under headerReadTimeoutMs (EX-30)
+        wrap streams: BufferedByteSource in, buffered OutputStream out, raw OutputStream rawOut
+        negotiated = negotiateProtocol(socket, in)   # ALPN or h2c preface
+        build ConnectionContext
+        dispatch to http1Protocol.run(ctx) or http2Protocol.run(ctx)
+    finally:
+        activeSockets.remove(socket); scratchPool.release(scratch)
+```
+
+`Http1Connection.run(ConnectionContext)` is where HTTP/1.1 semantics actually live: the
+keep-alive loop, the idle/header/body deadline transitions (Phase 1), the `MalformedRequestException`
+rejection path, the WebSocket upgrade handoff, and the response write.
+
+## `ConnectionScratch` and `ScratchPool` (`EX-06`)
+
+`ThreadLocal` is the right idiom when "one per thread" means "one per core" — a bounded
+platform-thread pool. Flash runs one **virtual** thread per connection
+(`Executors.newVirtualThreadPerTaskExecutor()`), so a `ThreadLocal` there means one per
+*connection*, with no upper bound: at 100 000 concurrent connections, an 8 KB relay buffer alone
+would be ~800 MB that a bounded pool would otherwise cap.
+
+`ConnectionScratch` is therefore an explicit, plain object (decimal-encoding buffer, streaming
+relay buffer, the WebSocket-handshake `MessageDigest`) acquired from a `ScratchPool` at
+connection start and released at connection end. The pool is a bounded *cache*, not a
+leak-free arena: above its bound (`min(availableProcessors * 64, 4096)` by default), a released
+scratch is simply dropped for the garbage collector rather than queued, so an unusually large
+burst of connections cannot grow it without limit.
+
+The routers use an explicit per-connection scratch passed through `AbstractRouter.route`; neither
+`FastPathRouterImpl` nor `FastPathWsRouterImpl` retains connection state in a `ThreadLocal`.
+
+## The `ConnectionProtocol` seam (R1 / `DEC-02`)
+
+```java
+public interface ConnectionProtocol {
+    void run(ConnectionContext ctx) throws IOException;
+}
+```
+
+`ConnectionRunner` decides h1 vs h2 exactly once, immediately after ALPN/preface detection, and
+dispatches to `Http1Connection` or `Http2Connection`. Neither implementation is aware the other
+exists — `dev.relism.flash.http1` and `dev.relism.flash.http2` do not import each other, enforced
+by `PackageBoundaryTest`.
+
+## Graceful shutdown (`EX-32`)
+
+Two stages, driven by `ServerLifecycle.stop()`:
+
+1. **Stop accepting.** Every listener socket is closed immediately; `stopped` flips to `true`.
+2. **Drain, then force-close.** `Http1Connection`'s request loop checks `ctx.stopped()` twice:
+   once before waiting for the next request (exits immediately if already stopped, rather than
+   waiting out the idle-keep-alive timeout), and again right before writing the *current*
+   response — forcing `Connection: close` on it even if the response's own `Connection` header
+   logic would have said keep-alive, and even if shutdown began *while the handler was running*
+   (the common case). `ServerLifecycle.stop()` polls `activeSockets` for up to
+   `shutdownDrainTimeoutMs`, then force-closes whatever remains and shuts down the executor.
+
+HTTP/2 shutdown sends the two-stage `GOAWAY` sequence from RFC 9113 §6.8 before the lifecycle's
+drain deadline force-closes remaining sockets.
+
+## What changed for WebSocket (`EX-01`, `EX-11`, `EX-12`, `EX-13`)
+
+- **`EX-01`**: `WebSocketSession`'s two blocking-write sites (`close`, `writeFrame`) now
+  serialize on a `ReentrantLock` instead of `synchronized (out)` — a virtual thread blocking
+  inside `synchronized` pins its carrier platform thread on Java 21 (JEP 491, which removes
+  this, is JDK 24+). `ReentrantLock` unmounts the blocked virtual thread instead.
+- **`EX-11`**: `readFrame` used to read the extended-length and mask-key bytes one at a time.
+  It now reads that whole variable-length remainder in a single bounded `readFully` into the
+  existing `hdrScratch` array, then decodes with shifts.
+- **`EX-12`**: `readFrame` now reassembles continuation frames into one logical message (bounded
+  by the same buffer a single frame already had), enforces the masking direction RFC 6455 §5.1
+  requires for this session's role, validates the opcode against the RFC's defined set, enforces
+  control-frame constraints (not fragmented, ≤125 bytes), and reports violations via
+  `WebSocketProtocolException` carrying the correct close code (1002 protocol error, 1009
+  message too big) for `WebSocketLoop` to send before closing.
+- **`EX-13`**: the `Connection` header is a comma-separated token list, not a single value —
+  `Http1KeepAlive.tokenListContains` is the one scanner both the keep-alive decision and
+  `WebSocketUpgrade`'s `Connection: Upgrade` check use, so they cannot drift apart again.

flash/docs/http2/BASELINES.md

@@ -0,0 +1,44 @@
+# HTTP performance baselines
+
+These numbers are regression controls, not cross-machine promises. They were measured on
+2026-08-13 under Linux 6.12/KVM, six exposed cores of an AMD Ryzen 7 1700X, Temurin 21.0.11 and
+JMH 1.37. CI uses short independent forks for allocation and sample latency so the sampling
+harness does not contaminate `gc.alloc.rate.norm`.
+
+## Gated hot paths
+
+| Benchmark | B/op | p50 ns | p99 ns | p999 ns | CI p99 ceiling ns |
+|---|---:|---:|---:|---:|---:|
+| h1 parse and route | 0.022 | 540 | 33,472 | 60,822 | 45,000 |
+| h2 pooled stream lifecycle | 0.010 | 530 | 2,138 | 37,724 | 2,900 |
+| h2 response encoding | 0.004 | 210 | 993 | 14,626 | 1,350 |
+| HPACK browser-request decode | 0.015 | 730 | 5,245 | 27,577 | 7,100 |
+| HPACK typical-response encode | 0.003 | 180 | 620 | 12,025 | 850 |
+| frame read/validate/discard | 0.006 | 70 | 1,999 | 90,508 | 2,700 |
+
+The sub-byte allocation values occur with no collection and are JMH/GC-profiler rate
+normalization noise. The CI allocation ceiling is 0.05 B/op. A benchmark exceeding it fails; a
+baseline or ceiling change requires an explicit edit and justification here.
+
+The table records the higher percentile observed across three consecutive controlled runs; this is
+important because short sample-mode runs on the shared KVM host showed visible scheduler noise.
+The p999 values expose those tails but are recorded rather than gated. The p99 ceilings are the
+worst observed p99 plus about 35% headroom.
+
+## HTTP/1 historical comparison
+
+The plan required a pre-Phase-1 number, but no benchmark was committed at that point. Phase 17
+reconstructed the current `RequestPipelineBenchmark.parseAndRoute` fixture against Phase 0 commit
+`db6e4a4` in a detached worktree and ran both revisions on the same host and JVM:
+
+| Revision | ns/op | B/op |
+|---|---:|---:|
+| Phase 0 (`db6e4a4`) | 976.195 ± 45.924 | 224.007 |
+| Phase 17 | 1,024.602 ± 50.744 | 0.007 |
+
+The hardened parser's mean is 5.0% higher and removes effectively all 224 B/op. The 99.9%
+confidence intervals overlap (`930.271–1,022.120` ns for Phase 0 and `973.858–1,075.345` ns for
+Phase 17), so this run does not establish a statistically significant latency regression. This is
+an honest reconstruction, not a claim that an absent historical run existed. Phase 17 recovered
+about 4.5% by having `RequestParser` populate `Http1HeaderMap`'s zero-copy index during the same
+validated header pass instead of rescanning every line; all security checks remain in that path.

flash/docs/http2/CLEARTEXT.md

@@ -0,0 +1,29 @@
+# HTTP/2 cleartext
+
+TLS HTTP/2 and cleartext HTTP/2 have independent rollout controls:
+
+- `http2Enabled` advertises `h2` through TLS ALPN.
+- `http2CleartextEnabled` accepts the HTTP/2 prior-knowledge preface on plaintext listeners.
+
+Both default to `false`. Cleartext support follows RFC 9113 prior knowledge. The obsolete
+HTTP/1.1 `Upgrade: h2c` transition is intentionally unsupported.
+
+## Header conversion
+
+`HopByHopHeaders` is the single policy used at connection boundaries. It removes fields named by
+`Connection`, the standard hop-by-hop set, HTTP/2-forbidden fields and pseudo-fields. `TE` is
+forwarded only as `trailers` when the target is HTTP/2. Tests execute the same policy for all four
+HTTP/1.1 and HTTP/2 source/target combinations.
+
+## Authority and 421
+
+On TLS HTTP/2 connections, Flash checks `:authority` against the selected certificate's DNS/IP
+subject alternative names. An authority outside that served set receives `421 Misdirected
+Request`, allowing a coalescing client to retry on a different connection. Exact names and
+single-label wildcards are supported; h2c has no certificate identity and is unaffected.
+
+An outbound HTTP/2 client and reverse-proxy adapter (`Http2Client`, `HttpProxy`) were built
+against this cleartext support but had no caller anywhere in `flash` core — an HTTP/1.1+2 server
+framework has no business shipping an outbound client. That code has been removed; if a
+reverse-proxy capability is needed later, it belongs in its own `flash-extensions/flash-ext-*`
+module, not in core.

flash/docs/http2/COMPLIANCE.md

@@ -0,0 +1,84 @@
+# HTTP/2 compliance
+
+This document records the repeatable protocol gate for Flash's HTTP/2 server. The automated
+matrix runs from Maven; external tools are selected through system properties so local builds
+without them skip only the corresponding interoperability adapter. CI installs and enables every
+command-line client listed below.
+
+## h2spec
+
+Validated on 2026-08-13 with h2spec 2.6.0.
+
+| Listener | Cases | Failures | Skips |
+|---|---:|---:|---:|
+| TLS with ALPN `h2` | 146 | 0 | 0 |
+| Cleartext prior knowledge on the mixed HTTP/1.1 + HTTP/2 port | 145 | 0 | 0 |
+
+`H2SpecComplianceTest` parses h2spec's JUnit XML and fails on a failure, error, or skipped case.
+The cleartext selection omits only `http2/3.5/2`, which sends a complete invalid HTTP/2 preface.
+That case assumes a dedicated HTTP/2 endpoint. Flash deliberately has one cleartext port that
+selects HTTP/2 only when the 24-byte prior-knowledge preface matches; any other initial bytes are
+HTTP/1.1 input. RFC 9113 section 3.3 defines the exact preface as the cleartext protocol selector,
+while section 3.4's `PROTOCOL_ERROR` applies after an endpoint is operating as HTTP/2. The HTTP/2
+state machine itself does return `GOAWAY(PROTOCOL_ERROR)` for a complete invalid preface, covered
+byte-for-byte by `invalid-preface.hex`. Excluding the mixed-port negotiation case therefore does
+not waive an HTTP/2 state-machine requirement.
+
+## Interoperability
+
+Automated results recorded on 2026-08-13:
+
+| Client | Version | Mode and coverage | Result |
+|---|---|---|---|
+| curl | 8.5.0, libnghttp2 1.59.0 | TLS and h2c; GET, POST, 2 MiB upload/download | pass |
+| Java `HttpClient` | Temurin 21.0.11+10 | TLS; GET, POST, large bodies and multiplexing | pass |
+| nghttp | nghttp2 1.59.0 | TLS and h2c; verbose SETTINGS/HEADERS/DATA trace, POST and 2 MiB download | pass |
+| grpcurl | 1.9.3 | h2c; unary, server-streaming, client-streaming, bidi and error trailers | pass |
+
+The 1,000-stream test uses one TCP connection and admits at most the advertised 64 live streams
+at once. This tests 1,000 multiplexed stream lifecycles without contradicting
+`SETTINGS_MAX_CONCURRENT_STREAMS` or weakening the production memory bound.
+
+Chrome and Firefox are a release smoke test rather than a CI dependency. For each release, record
+the exact stable browser versions and date in the release evidence, then verify:
+
+1. Load a TLS route and confirm `h2` in the browser network protocol column.
+2. Exercise GET, POST, a large upload and a large streamed download.
+3. Open the same registered WebSocket route over HTTP/1.1 and RFC 8441, exchange a fragmented
+   message larger than one flow-control window, and close from each side once.
+4. Confirm no certificate, console, failed-request, or retry-to-HTTP/1.1 warnings.
+
+This manual row is intentionally not represented as an automated pass: browser release testing
+must record the browsers actually shipped at release time rather than a stale development image.
+
+## Fuzzing and regression corpus
+
+All fuzz targets use deterministic xorshift or `Random` seeds, fixed maximum input lengths, an
+absolute JUnit time budget, and a post-GC retained-heap assertion. Untyped runtime failures fail
+the test immediately. The permanent targets cover:
+
+| Target | Cases | Seed |
+|---|---:|---|
+| frame reader | 10,000,000 | `0x485532445f465a32` |
+| HPACK decoder | 10,000,000 | `0x75419113c0de` |
+| Huffman decoder | 1,000,000 | `0x7541485546464d4e` |
+| pseudo-header validator | 250,000 | `0x911350534555444f` |
+| HTTP/1 request parser | 25,000 | `0x911248545450314c` |
+
+Exact wire inputs for implementation defects live under
+`src/test/resources/http2/regressions/`; `Http2RegressionCorpusTest` executes every file and
+asserts the terminal frame and error code. The nightly `Http2SoakTest` defaults to ten minutes of
+GET, POST, streaming DATA, reset and PING traffic, with retained-heap assertions. A short run can
+be requested with `-Dflash.http2.soak=true -Dflash.http2.soak.seconds=10`.
+
+## Deliberately absent features
+
+- HTTP/2 server push is not exposed. A client cannot send `PUSH_PROMISE` to a server (RFC 9113
+  section 6.6); receiving one is a connection error. Flash does not originate push.
+- RFC 7540 dependency-tree priority scheduling is not implemented. RFC 9113 section 5.3.2
+  deprecates the scheme; PRIORITY frames are validated and ignored as required.
+- `Upgrade: h2c` is not implemented. RFC 9113 section 3.1 removed the HTTP/1.1 upgrade mechanism;
+  cleartext support uses section 3.3 prior knowledge.
+
+These omissions do not create alternate request/response APIs: HTTP/1.1 and HTTP/2 remain peers
+behind the transport protocol boundary.

flash/docs/http2/CONNECTION.md

@@ -0,0 +1,70 @@
+# HTTP/2 connection control
+
+`Http2Connection` owns only connection-level protocol state. It verifies the preface, drives the
+frame reader, dispatches control frames and performs shutdown. HPACK fragment extraction and decode
+live in `Http2HeaderBlockDecoder`; socket serialization remains exclusively in
+`Http2FrameWriter`. Stream dispatch and application handlers are separate layers.
+
+Each accepted HTTP/2 socket receives a new `Http2Connection`. Sharing the stateless
+`Http1Connection` implementation is safe, but sharing an HTTP/2 instance would leak dynamic HPACK,
+SETTINGS, flow-control and GOAWAY state between peers.
+
+## Demultiplexing invariant
+
+The demux thread never invokes application work. It reads and validates frames, updates bounded
+connection state, and enqueues or directly writes control frames. A registered handler cannot delay
+SETTINGS or PING processing. The connection reader polls at a short interval so server shutdown is
+observed promptly, while `Http2FrameReader` retains one non-renewable absolute deadline for a
+partially received frame; polling therefore does not weaken slow-frame protection.
+
+## Settings
+
+| Identifier | Default | Validation and handling |
+|---|---:|---|
+| `HEADER_TABLE_SIZE` | 4096 | Unsigned 32-bit; locally capped |
+| `ENABLE_PUSH` | 1 | Only 0 or 1; Flash advertises 0 |
+| `MAX_CONCURRENT_STREAMS` | unlimited | Unsigned 32-bit |
+| `INITIAL_WINDOW_SIZE` | 65535 | At most 2^31-1 |
+| `MAX_FRAME_SIZE` | 16384 | 16384 through 16777215 |
+| `MAX_HEADER_LIST_SIZE` | unlimited | Unsigned 32-bit |
+
+Unknown identifiers are ignored. A payload is validated as a transaction before values are
+committed. The initial-window delta is handed to the stream table as one operation: negative stream
+windows are valid, but any result above 2^31-1 rejects the complete update with
+`FLOW_CONTROL_ERROR`. Every non-ACK SETTINGS frame receives an empty ACK; locally sent settings are
+bounded and have an acknowledgement deadline.
+
+## Priority control writes
+
+`Http2FrameWriter` has one priority MPSC lane in front of its ordinary stream-data lane. PING and
+SETTINGS acknowledgements, RST_STREAM and GOAWAY use reusable control intents from the connection
+scratch. They can overtake queued DATA but never split or interrupt a socket write already in
+progress. Both PING and SETTINGS response queues are bounded.
+
+## Shutdown
+
+Graceful shutdown follows the two-stage protocol:
+
+1. Send GOAWAY with last-stream-id 2^31-1 and `NO_ERROR`.
+2. Send a connection PING and wait for its matching ACK, establishing a round trip.
+3. Send a second GOAWAY with the real last processed stream id, then close after current work.
+
+A connection error instead sends one GOAWAY with the precise error code, the real last processed
+stream id and a bounded diagnostic string. A preface mismatch closes silently because the peer has
+not established a valid HTTP/2 connection.
+
+## Verification
+
+The reusable control lifecycle (preface, SETTINGS/ACK, PING/PONG, WINDOW_UPDATE and received
+GOAWAY) measures 974.263 ns/op and 0.008 B/op on JDK 21.0.11; the allocation figure is the JMH GC
+profiler noise floor with no collections. `curl 8.5.0` using h2c prior knowledge completed the
+handshake and observed both clean GOAWAY stages. It exits with code 56 because this phase
+deliberately sends no response HEADERS or DATA; those arrive with the response and stream phases.
+
+h2spec 2.6.0 passes 28 of the 35 selected section 3, 4, 6.5, 6.7, 6.8 and 6.9 cases, including all
+connection-owned SETTINGS validation, PING, GOAWAY, frame-format, HPACK interleaving and
+connection-window cases. Six failures require response HEADERS/DATA or per-stream flow control and
+remain assigned to the response, stream and DATA phases. The seventh is h2spec's expectation of a
+GOAWAY after an invalid preface; Flash intentionally closes without writing because no valid HTTP/2
+connection exists yet, as permitted by RFC 7540 §3.5 and required by this implementation's preface
+contract.

flash/docs/http2/FLOW-CONTROL.md

@@ -0,0 +1,48 @@
+# HTTP/2 bodies and flow control
+
+HTTP/2 applies flow control independently to the connection and to every stream. Flash advertises
+a 1 MiB receive window at both levels and sends WINDOW_UPDATE only after the application has
+consumed at least half a window. A DATA frame decrements both windows by its complete payload
+length, including the pad-length byte and padding; only its unpadded data reaches the handler.
+
+## Request bodies
+
+Known bodies up to 64 KiB remain in one reusable contiguous stream buffer. Their handler is
+dispatched at END_STREAM, and `RequestBody.bytes()` performs the only allocation: the byte array
+returned to application code. For a 1,024-byte body JMH reports exactly 1,040 B/op, the array plus
+its object header, with no framework allocation around it.
+
+Larger or unknown-length bodies dispatch after request headers. DATA is copied out of the frame
+reader into a connection-owned pool of 64 reusable 16 KiB buffers. Small adjacent frames coalesce
+inside a buffer, so the pool is bounded by bytes rather than frame count. The existing
+`RequestBody.stream()` blocks only the handler's virtual thread when data is absent. Buffers return
+to the pool as reads consume them, and that consumption reopens both receive windows. If a handler
+does not read its body, the normal post-handler drain performs the same bounded consumption.
+
+The connection window and pool both cover exactly 1 MiB, so the peer can never hold more credit
+than the server can store before backpressure takes effect. Per-stream accepted body bytes remain
+bounded by `MAX_REQUEST_BODY_SIZE`. Declared content length is parsed without a String and checked
+against the unpadded DATA total at END_STREAM.
+
+## Responses
+
+Fixed byte arrays, known-length streams and unknown-length streams all use one resumable
+`Http2ResponseWriter`. It emits DATA frames no larger than the peer's frame limit, the available
+connection window, the available stream window and the reusable 16 KiB relay buffer. A
+WINDOW_UPDATE schedules the stream on the shared virtual-thread executor; application streams are
+never read by the demultiplexer.
+
+`Response.chunked(InputStream)` means unknown-length streaming at the application API. HTTP/2 has
+no chunked transfer coding, so Flash emits ordinary DATA followed by END_STREAM and never sends a
+`transfer-encoding` field. `Response.stream(InputStream, length)` emits `content-length` and fails
+the stream if the source ends before that length.
+
+## Verification
+
+- A real Java HTTP/2 client uploads and downloads 100 MiB over TLS; both directions are validated
+  byte-for-byte without materializing the test payload.
+- A synthetic 100 MiB response proves serialized scratch storage stays below 64 KiB.
+- h2spec sections 5, 6.1, 6.9 and 8: 50 passed, one h2spec-skipped case, zero failures.
+- Clean Maven build with JMH sources: 633 tests, no failures.
+- JMH request streaming: 159.408 ns/op, 0.001 B/op, no GC.
+- JMH response streaming frame: 219.090 ns/op, 0.002 B/op, no GC.

flash/docs/http2/FRAMES.md

@@ -0,0 +1,147 @@
+# The frame layer
+
+Audience: contributors. This is the design record for `dev.relism.flash.http2.frame`'s frame
+reading, validation, and writing — the 9-byte header and payload boundary, with no connection
+semantics, no streams, and no HPACK above it.
+
+## Why this is simpler than the h1 parser
+
+HTTP/1.1 request parsing must scan for `\r\n\r\n` (`RequestParser`, `ByteScan.indexOfCrLfCrLf`)
+because nothing in the h1 wire format states the header block's length up front. HTTP/2 states
+every frame's payload length in the first three bytes of its 9-byte header — nothing is ever
+scanned for. `Http2FrameReader` is a length-prefixed reader and nothing more: read 9 bytes,
+decode the length, ensure that many more bytes are available, done.
+
+## The wire format
+
+```
++-----------------------------------------------+
+|                 Length (24)                   |
++---------------+---------------+---------------+
+|   Type (8)    |   Flags (8)   |
++-+-------------+---------------+-------------------------------+
+|R|                 Stream Identifier (31)                      |
++=+=============================================================+
+|                   Frame Payload (0...)                      ...
++---------------------------------------------------------------+
+```
+
+`R` (RFC 9113 §4.1) is reserved and MUST be ignored on receipt — `FrameHeader.reset` masks it
+out of `streamId()` once, so no caller has to remember to.
+
+## Package layout
+
+```
+dev.relism.flash.http2.frame
+├── FrameType            the 10 known types + per-type validation descriptor (min/max length, stream-id rule)
+├── FrameFlags            END_STREAM/ACK/END_HEADERS/PADDED/PRIORITY bit constants + predicates
+├── FrameHeader           flyweight over a read buffer: length/type/flags/streamId/payloadOffset
+├── Http2FrameReader      length-prefixed reader, RequestParser's buffer/compaction discipline
+├── FrameValidator        table-driven per-type RFC validation, specific error code per rule
+├── Padding               RFC 9113 §6.1/§6.2 pad-length byte + trailing padding, DATA/HEADERS
+├── FrameWriteBuffer      beginFrame()/endFrame() length back-patching over a ByteWriter
+├── Http2FrameWriter      the connection's single serialized writer
+├── WriteIntent           caller-owned serialized frame batch
+└── IntrusiveMpscQueue    allocation-free contended-write queue
+```
+
+## The validation table
+
+Every rule below is enforced by `FrameValidator.validate(FrameHeader, insideHeaderBlock)`, in
+this order: unknown-type handling, `SETTINGS`' modulus-6 special case, the generic
+min/max length bounds, the `MAX_FRAME_SIZE_LOCAL` ceiling, the stream-id rule, then
+`PUSH_PROMISE`'s always-reject rule.
+
+| Type | Code | Length | Stream id | Notes / RFC |
+|---|---|---|---|---|
+| DATA | 0x0 | 0..MAX_FRAME_SIZE | required (≠0) | §6.1. Padding via `Padding.unpad`. |
+| HEADERS | 0x1 | 0..MAX_FRAME_SIZE | required (≠0) | §6.2. Padding and optional PRIORITY fields are parsed before HPACK. |
+| PRIORITY | 0x2 | exactly 5 | required (≠0) | §6.3. Deprecated (§5.3.2) — parsed, discarded, never acted on. |
+| RST_STREAM | 0x3 | exactly 4 | required (≠0) | §6.4. The 4 bytes are the error code. |
+| SETTINGS | 0x4 | multiple of 6 | forbidden (=0) | §6.5. Modulus checked before the generic bounds. |
+| PUSH_PROMISE | 0x5 | ≥4 | required (≠0) | §6.6. Always `PROTOCOL_ERROR` from a client — never sent by Flash. |
+| PING | 0x6 | exactly 8 | forbidden (=0) | §6.7. Opaque 8-byte payload, echoed on ACK. |
+| GOAWAY | 0x7 | ≥8 | forbidden (=0) | §6.8. Last-stream-id (4) + error code (4) + optional debug data. |
+| WINDOW_UPDATE | 0x8 | exactly 4 | either | §6.9. 0 = connection window, ≠0 = one stream's window. |
+| CONTINUATION | 0x9 | 0..MAX_FRAME_SIZE | required (≠0) | §6.10. Continues a header block; see the flood guard below. |
+| *(unrecognised)* | >0x9 | — | — | §4.1: ignored outside a header block, `PROTOCOL_ERROR` inside one (§6.10). |
+
+**The error code is not uniform per type** — a `SETTINGS` frame with a bad length is
+`FRAME_SIZE_ERROR`; the same frame with a non-zero stream id is `PROTOCOL_ERROR`. Every violation
+in the table above carries its own RFC citation and the specific code that citation mandates;
+`FrameValidatorTest` has one test per row asserting the exact code, not merely "an exception".
+
+## Ignore vs. reject policy
+
+RFC 9113 §4.1 makes unknown frame types part of the protocol's extension mechanism: an endpoint
+that does not recognise a type MUST read and discard its payload, never reject the connection for
+it. `FrameType.fromCode` returns `null` for anything above `CONTINUATION` (0x9); `FrameHeader`
+still exposes the raw `typeCode()` for logging even when `type()` is `null`.
+
+The one exception (§6.10): if an unrecognised-type frame arrives **between** a HEADERS/
+PUSH_PROMISE frame that lacked `END_HEADERS` and the CONTINUATION that eventually sets it, the
+HPACK decoder's state has nowhere to put that frame's bytes without desynchronizing — so this one
+case *is* a `PROTOCOL_ERROR`, tracked by `FrameValidator.validate`'s `insideHeaderBlock`
+parameter (owned and threaded through by the connection loop, which is the only caller
+that knows whether a header block is currently open).
+
+`PRIORITY` frames are a different kind of "ignore": they are a recognised, well-formed type that
+Flash chooses not to act on (RFC 9113 §5.3.2 deprecates priority signalling and permits an
+implementation to disregard it) — they are still fully parsed and validated like any other frame,
+just never influence scheduling. `PUSH_PROMISE` is the opposite: recognised, but **always**
+rejected when received (Flash advertises `SETTINGS_ENABLE_PUSH=0` and never sends one itself), so
+receiving one at all can only mean the peer has the client/server roles backwards.
+
+## Buffer discipline and the frame-size defence
+
+`Http2FrameReader` never grows its buffer to accommodate a declared length before checking that
+length against `Http2Limits.MAX_FRAME_SIZE_LOCAL` — the check happens first, so a hostile 16 MB
+declared length is rejected at the cost of reading 9 bytes, not at the cost of a 16 MB
+allocation. This mirrors `RequestParser`'s own `EX-08` discipline (bound the request line before
+trusting it) applied to the frame layer's own attack surface.
+
+The buffer itself follows `RequestParser`'s compact-before-grow policy: unconsumed bytes slide to
+offset 0 when there is room to do so without growing, and growth only happens when compaction
+alone cannot make room — bounded, because the reader's own length check already rejected
+anything that would require growing past `9 + MAX_FRAME_SIZE_LOCAL`.
+
+## Padding
+
+`Padding.unpad` locates the actual data range within a `PADDED` frame's payload: 1 byte of
+pad-length, then data, then that many padding bytes (whose contents carry no meaning — they exist
+only to obscure payload size from network observers). A pad length greater than or equal to the
+whole payload length is `PROTOCOL_ERROR` (RFC 9113 §6.1), checked before any arithmetic that
+could otherwise underflow. Flow-control accounting for padded DATA frames (RFC 9113 §6.9.1: the
+*whole* payload counts against the window, not just the data) is applied by
+`Http2FlowController`; `Padding` only locates the data range.
+
+## Writing: `FrameWriteBuffer`'s back-patching
+
+A frame's length is rarely known before its payload is serialized (an HPACK-encoded header block,
+in particular, has no cheap way to be measured in advance). `FrameWriteBuffer.beginFrame` writes
+a 9-byte header with a placeholder length; the caller writes the payload directly through the
+same `ByteWriter`; `endFrame` computes the actual length from how far the writer has advanced and
+rewrites the three length bytes in place. This is *why* `Http2FrameWriter` serializes a
+complete buffer before ever taking the connection lock, rather than streaming bytes as they are
+produced — streaming would need the length upfront, which back-patching deliberately avoids
+needing.
+
+## Buffered-source deadline regression
+
+`BufferedByteSource`'s deadline mechanism (`EX-07`'s actual fix) turned out to have zero dedicated
+unit tests and an unconditional `socket.setSoTimeout(...)` call that NPE'd against the `null`
+socket every isolated unit test in this codebase uses. Found while writing
+`Http2FrameReaderTest`, fixed, and given its own regression suite (`BufferedByteSourceTest`).
+
+## Testing
+
+- `Http2FrameReaderTest` — round-trips every frame type, boundary lengths (0, 1, 16383, 16384,
+  16385), a frame split across three socket reads, a frame exactly filling the initial buffer,
+  multiple sequential frames, clean-EOF-vs-mid-frame-EOF, and reserved-bit masking.
+- `FrameValidatorTest` — one test per RFC-mandated rejection above, asserting the specific
+  `Http2ErrorCode`.
+- `Http2FrameReaderFuzzTest` — 10 000 000 random-length (0–64 byte), random-content inputs; only
+  `Http2Exception`, `EOFException`, or `SocketTimeoutException` may escape. Green, ~14s.
+- `PaddingTest` — every boundary of the pad-length arithmetic, including the exact
+  `padLength == payloadLength - 1` (maximum valid) and `padLength >= payloadLength` (rejected)
+  cases.

flash/docs/http2/HPACK.md

@@ -0,0 +1,82 @@
+# HPACK decoder
+
+This document records the implementation constraints of Flash's RFC 7541 decoder. It is
+contributor documentation, not application API documentation.
+
+## Representation model
+
+`HpackDecoder` accepts all RFC 7541 field representations: indexed fields, literals with
+incremental indexing, literals without indexing, never-indexed literals, and dynamic-table size
+updates. Prefix integers are bounded against overflow, Huffman padding and EOS are validated, and
+decoded string lengths are checked while bytes are produced.
+
+The static table is stored as 61 immutable name/value byte pairs. Encoder-oriented reverse lookup
+uses fixed open-addressed integer tables built during class initialization; lookup never converts
+header bytes to `String` and never calls `HashMap` on the hot path.
+
+The dynamic table owns a bounded byte arena and a ring of primitive entry descriptors. Entry size
+is `name length + value length + 32`, and eviction is oldest-first as required by RFC 7541 §4.1.
+When the arena tail is too short, live entries are compacted into a contiguous prefix. This avoids
+segmented views in every consumer and keeps indexed fields cheap to copy.
+
+## Ownership and eviction safety
+
+Views emitted by `HeaderSink` are callback-scoped. Production decoding targets a reusable
+`HpackHeaderBlock` owned by the stream, which copies each name and value into its own arena.
+
+The copy is required for correctness. Consider stream A referencing a dynamic-table entry while
+its handler is running. The connection thread can then decode stream B, evict that entry, and
+reuse its bytes. If stream A retained the dynamic-table view, its headers would silently change.
+Per-stream storage removes that race without reference counting or synchronization.
+
+The precise copy model is:
+
+- HTTP/1.1 copies nothing per request but scans header bytes in the connection buffer.
+- HTTP/2 copies decoded request headers into stream-owned storage because multiplexed handlers
+  outlive subsequent HPACK mutations.
+- Novel incrementally-indexed fields are also copied once into the connection's dynamic table.
+
+`HpackEvictionRaceTest` contains both the unsafe borrowed-view demonstration and the stable
+stream-owned result.
+
+## Header-list rejection
+
+The decoder counts RFC header-list size cumulatively. Once the configured limit is crossed it
+stops emitting fields, but continues parsing the entire block and applying dynamic-table updates.
+Only after the block ends does it throw `HeaderListSizeException`. The stream layer can reject the
+request while the connection's compression state remains synchronized.
+
+## CONTINUATION assembly
+
+`ContinuationAssembler` copies HEADERS and CONTINUATION fragments into one bounded connection
+buffer. It rejects interleaving, stream-id changes, excessive continuation count, and blocks that
+exceed the configured capacity. `SegmentedByteView` is intentionally not used here: RFC 9113 §6.10
+requires a contiguous, non-interleaved continuation sequence, and one bounded copy makes the HPACK
+decoder and all downstream views simpler.
+
+## Verification
+
+- RFC 7541 Appendix C.1–C.6 vectors, including dynamic-table state after every sequence.
+- Invalid integer, Huffman, index, size-update, and header-list inputs.
+- Ten million deterministic random blocks; only typed protocol rejections may escape.
+- JMH `-prof gc`: `decodeStaticRequest` measured 102.725 ns/op and 0.001 B/op on JDK 21.0.11. The
+  latter is the profiler's sampling noise floor; no garbage collections occurred.
+
+## Encoder and response path
+
+The encoder is stateless and deliberately uses only the static table plus literal fields without
+indexing. It emits a dynamic-table-size update of zero at the start of the connection's first
+response block. This avoids mutable compression state shared by concurrent streams; the trade-off
+is a few more wire bytes for repeated custom response fields.
+
+Status and known content-type fields are HPACK-encoded during class initialization. The cached Date
+header refreshes both its HTTP/1 and HPACK forms once per second. Runtime values are raw literals by
+default; `FlashConfiguration.h2HuffmanDynamicValues` enables Huffman coding when deployment-specific
+measurements justify its CPU/wire-size trade-off.
+
+`Http2ResponseWriter` is reusable per stream. It lowercases field names, removes forbidden
+connection-specific fields, enforces the peer's header-list bound, keeps HEADERS and CONTINUATION
+frames in one write intent, and appends a small fixed DATA body when flow-control permits.
+
+JMH `-prof gc` measured the representative response path at 174.309 ns/op and 0.001 B/op on JDK
+21.0.11, with no garbage collections. The reported allocation is the profiler noise floor.

flash/docs/http2/PERFORMANCE.md

@@ -0,0 +1,99 @@
+# HTTP/2 performance
+
+## Method
+
+Measurements were taken on 2026-08-13 under Linux 6.12/KVM with six exposed AMD Ryzen 7 1700X
+cores, Temurin 21.0.11, JMH 1.37 and nghttp2 1.59.0. JMH component benchmarks use prepared,
+reusable protocol state and forked JVMs. `h2load` exercises the real cleartext server on loopback;
+Flash and nghttpd run on the same host in alternating order. Results are snapshots, not promises
+for different hardware.
+
+No "unmatched throughput" claim is supported. nghttpd is normally faster in this matrix; Flash's
+numbers include framework routing, request-model assembly and handler dispatch that the static
+reference server does not.
+
+## Component results
+
+The CI-controlled allocation and percentile numbers are in `BASELINES.md`. Additional average
+time measurements from the same run were:
+
+| Scenario | Result |
+|---|---:|
+| h2 responses across 1 live stream | 74.405 ns |
+| h2 responses across 8 live streams | 705.368 ns |
+| h2 responses across 64 live streams | 5,591.368 ns |
+| h2 responses across 256 live streams | 28,961.681 ns |
+| h2 POST lifecycle with 1 KiB DATA | 741.031 ns, 0.005 B/op |
+| 1 MiB streaming response | 78,681.815 ns |
+
+The multiplexing benchmark reports one complete response-encoding pass across all live streams,
+not per-stream time. `Http2BodyBenchmark` separately covers the 1 KiB request-body shape and the
+1 MiB response shape. `FrameWriterBenchmark` retains the Phase 3 contention matrix and its
+per-write latency distribution.
+
+## End-to-end h2load comparison
+
+Each row uses at least 1,000 requests. Requested stream concurrency is capped first by Flash's
+advertised 64-stream setting and then to 4,096 aggregate active streams so the 1,000-connection
+rows remain bounded. Both requested and effective values are shown.
+
+| Connections | Requested/effective streams | Flash req/s | nghttpd req/s |
+|---:|---:|---:|---:|
+| 1 | 1 / 1 | 2,136.18 | 12,786.42 |
+| 1 | 10 / 10 | 18,396.56 | 83,521.26 |
+| 1 | 100 / 64 | 17,039.55 | 66,746.76 |
+| 10 | 1 / 1 | 11,247.08 | 37,838.66 |
+| 10 | 10 / 10 | 25,055.12 | 104,964.84 |
+| 10 | 100 / 64 | 3,878.28 | 67,303.81 |
+| 100 | 1 / 1 | 5,517.94 | 42,319.09 |
+| 100 | 10 / 10 | 1,818.52 | 26,732.25 |
+| 100 | 100 / 40 | 7,042.85 | 136,585.90 |
+| 1,000 | 1 / 1 | 1,393.17 | 3,877.62 |
+| 1,000 | 10 / 4 | 10,374.83 | 40,976.05 |
+| 1,000 | 100 / 4 | 49,622.10 | 85,344.40 |
+
+The matrix found a correctness issue before it produced these final numbers: closed streams still
+occupied live admission slots while their final write callback was pending. The bounded detach
+fix is recorded as EX-57 and covered by regression tests.
+
+## Tuning decisions
+
+| Knob | Measurement | Decision |
+|---|---|---|
+| 16 KiB / 64 KiB / 1 MiB response frame | 1 MiB stream: 104,071 / 97,996 / 98,769 ns in the non-Huffman sweep | Keep 16 KiB. The roughly 6% gain at 64 KiB does not justify 4x per-connection buffer exposure on this noisy host. |
+| 1 MiB initial receive window | 100 MiB Phase 11 transfer and the load matrix complete without flow stalls | Keep; it matches bounded receive capacity and changing it independently would not isolate a throughput claim. |
+| half-window WINDOW_UPDATE hysteresis | 1 MiB streaming and 100 MiB transfer complete with steady pooled reads | Keep; no per-frame update traffic and no demonstrated reason to weaken backpressure. |
+| 64 KiB inline body | 1 KiB inline materialization is one 1,040 B allocation; streaming steady state is ≈0 B/op | Keep the explicit one-array small-body tradeoff and stream larger bodies. |
+| 64 × 16 KiB DATA buffers | 1 MiB streaming is 78,682 ns with ≈0 B/op; h2load stays bounded | Keep; larger chunks did not produce a clear win beyond the frame-size sweep. |
+| `ScratchPool` bound | 64 objects per exposed CPU, capped at 4,096; full 1,000-connection matrix completes | Keep the capacity bound; it affects retained burst memory, not steady-state request instructions. |
+| word-at-a-time route compare | 14.289 ns versus 22.313 ns bytewise, 36.0% faster | Keep. |
+| SWAR header-end scan | 89.919 ns versus 128.460 ns scalar, 30.0% faster | Keep. |
+| `SlicePool` size 4 | Header/path/query view benchmarks remain allocation-free | Keep; size changes lifetime capacity, not lookup work, and four simultaneous borrowed views match the documented contract. |
+| runtime-value Huffman | representative response headers: 375.761 ns versus 180.129 ns at 16 KiB | Keep disabled by default; this header set is 109% slower to encode. |
+
+The frame-size/Huffman factorial produced counterintuitive variation in the body-only rows, so it
+was not used to claim a Huffman body effect: Huffman only prepares headers. This is treated as
+host noise rather than reverse-engineered into a preferred result.
+
+## Profiling
+
+async-profiler 4.4 was run against the representative browser HPACK decode. The top CPU leaves
+were `Huffman.decode` (72.67%), `HpackHeaderBlock.accept` (7.00%), `HpackDecoder.decode` (5.00%),
+JVM byte-array copy (4.00%), `HpackHeaderBlock.copy` (4.00%), `HpackStaticTable.name` (2.00%),
+`PooledSlice.reset` (1.00%), `PooledSlice.array` (0.67%), JVM byte-arraycopy (0.67%), and
+`HpackDecoder.decodeString` (0.67%). Each belongs to decoding, bounded arena ownership, or the
+copy that makes header lifetime independent of dynamic-table eviction; none is incidental
+locking or logging.
+
+The allocation profile produced no samples on the gated decode path. The realistic eight-writer
+lock profile produced no sampled contended locks; the writer benchmark measured 185.605 bursts/s,
+p50 1.2 µs, p99 5.3 µs and p999 41.6 µs. CPU, allocation and lock artifacts were generated under
+`/tmp/phase17-*` and are intentionally not committed.
+
+CI runs the complete suite with `-Djdk.tracePinnedThreads=full`. The forked allocation and p99
+gates run only under the Maven `jmh` profile; the h2load comparison remains informational and
+conditional because cross-runner throughput is not a stable correctness gate.
+
+The reconstructed Phase-0 HTTP/1 comparison is documented in `BASELINES.md`. Its confidence
+interval overlaps the Phase-17 result, while normalized allocation falls from 224.007 B/op to
+0.007 B/op.

flash/docs/http2/README.md

@@ -0,0 +1,43 @@
+# HTTP/2 in Flash
+
+Flash treats HTTP/1.1 and HTTP/2 as peer transports behind one connection boundary. TLS ALPN or
+the cleartext prior-knowledge preface selects a protocol once; both paths then feed the same
+router, `Request`, `Response`, handler, trailer, streaming and WebSocket APIs. HTTP/2 adds a
+bounded frame decoder, HPACK codec, stream state machine, two-level flow control and one serialized
+writer per connection. Application code does not branch on the wire protocol.
+
+The implementation is deliberately layered:
+
+```text
+listener / TLS
+    -> protocol negotiation
+        -> HTTP/1.1 parser ---------+
+        -> HTTP/2 frames + HPACK ----+-> shared request model -> router -> handler
+                                         shared response model
+        <- HTTP/1.1 serializer -----+
+        <- HTTP/2 stream writer ----+
+```
+
+This page covers the HTTP/2-specific layers only. The transport, message model, and byte
+primitives shared with HTTP/1.1 live in [`../core/`](../core/README.md).
+
+## Protocol layers
+
+- [Connection](CONNECTION.md) and [streams](STREAMS.md) — HTTP/2 connection and stream state.
+- [Flow control](FLOW-CONTROL.md) — request backpressure and streamed responses.
+- [Cleartext](CLEARTEXT.md) — prior knowledge and the 421 misdirected-request rule.
+- [WebSockets](WEBSOCKET.md) — RFC 8441 extended CONNECT using the existing WebSocket API.
+
+## Wire internals
+
+- [Serialized writer](WRITER.md) — the single-owner output path and contention model.
+- [Frames](FRAMES.md) — frame parsing, validation and error scope.
+- [HPACK](HPACK.md) — integer/Huffman coding and static/dynamic table ownership.
+
+## Operate and verify
+
+- [Security](SECURITY.md) — every HTTP/2 limit, default and abuse control.
+- [Troubleshooting](TROUBLESHOOTING.md) — GOAWAY/RST_STREAM diagnosis and protocol tracing.
+- [Compliance](COMPLIANCE.md) — h2spec, interoperability, fuzzing and deliberate omissions.
+- [Performance](PERFORMANCE.md) and [CI baselines](BASELINES.md) — measurements and regression
+  gates, including the comparison with nghttpd.

flash/docs/http2/SECURITY.md

@@ -0,0 +1,42 @@
+# HTTP/2 security controls
+
+HTTP/2 multiplexing lets one connection create disproportionate parser, stream and response work.
+Flash therefore combines structural bounds, flow-control bounds and rate bounds. Rate counters use
+two fixed half-window buckets, allocate nothing per frame and need no timer thread.
+JMH on JDK 21 measures one rate-counter increment at 38.083 ns/op and approximately
+`10^-4 B/op` (allocation noise floor, no GC).
+
+| Limit | Default | Defence / tuning guidance |
+|---|---:|---|
+| `MAX_CONCURRENT_STREAMS` | 64 | Bounds simultaneously retained stream state. |
+| `MAX_STREAMS_CREATED_PER_INTERVAL` | 400 / 10 s | Companion to Rapid Reset; tune with `h2MaxStreamsCreatedPerInterval`. |
+| `MAX_RESET_STREAMS_PER_INTERVAL` | 200 / 10 s | CVE-2023-44487 Rapid Reset; tune with `h2MaxResetStreamsPerInterval`. |
+| `MAX_CONTINUATION_FRAMES_PER_BLOCK` | 8 | CVE-2024-27316 CONTINUATION flood. |
+| `MAX_HEADER_LIST_SIZE` | 32 KiB | Stops HPACK expansion before fields reach stream storage. |
+| `MAX_HPACK_STRING_LENGTH` | 8 KiB | Bounds one decoded literal, including Huffman expansion. |
+| `MAX_SETTINGS_PER_INTERVAL` | 100 / 10 s | Bounds mandatory SETTINGS acknowledgements. |
+| `MAX_PINGS_PER_INTERVAL` | 200 / 10 s | Bounds mandatory PING acknowledgements. |
+| `MAX_USELESS_FRAMES_PER_INTERVAL` | 10,000 / 10 s | Aggregate CPU bound for PRIORITY, WINDOW_UPDATE, empty DATA and unknown frames. |
+| `MAX_SETTINGS_ACK_QUEUE_DEPTH` | 64 | Bounds queued SETTINGS control writes. |
+| `MAX_PING_QUEUE_DEPTH` | 64 | Bounds queued PING control writes. |
+| `MAX_EMPTY_DATA_FRAMES_PER_STREAM` | 1,000 | Stops DATA work that spends no flow-control credit. |
+| `INITIAL_WINDOW_SIZE_LOCAL` | 1 MiB | Matches the bounded DATA pool; consumption, not receipt, returns credit. |
+| `MAX_REQUEST_BODY_SIZE` | 100 MiB | Hard per-stream request body bound. |
+| `HEADER_BLOCK_ASSEMBLY_TIMEOUT_MS` | 10 s | Absolute HEADERS-to-END_HEADERS deadline. |
+| `STREAM_IDLE_TIMEOUT_MS` | 60 s | Cancels retained inactive streams; tune with `h2StreamIdleTimeoutMs`. |
+| `FRAME_READ_TIMEOUT_MS` | 20 s | Absolute partial-frame deadline. |
+| `WRITE_TIMEOUT_MS` | 30 s | Interrupts a socket writer blocked by a peer that stopped reading. |
+| `MAX_STREAMS_PER_CONNECTION` | 100,000 | Optional connection churn budget; zero disables, tune with `h2MaxStreamsPerConnection`. |
+| `MAX_BYTES_PER_CONNECTION` | disabled | Optional wire-byte budget; tune with `h2MaxBytesPerConnection`. |
+| `MAX_CONNECTION_LIFETIME_MS` | disabled | Optional lifetime rotation; tune with `h2MaxConnectionLifetimeMs`. |
+
+`h2AbuseRateIntervalMs` changes the rolling interval for reset and stream-creation operator
+limits. Breaching a connection-wide rate or budget produces GOAWAY `ENHANCE_YOUR_CALM`; malformed
+stream-local messages use the RFC-defined stream error. The ordinary write queue is bounded by the
+64 live streams and their single-in-flight response intent; control writes use the fixed scratch
+slots above, so a slow reader cannot create an unbounded application queue.
+
+The security suite covers Rapid Reset, stream churn, SETTINGS/PING/non-progress floods, 100,000
+CONTINUATION frames, HPACK expansion, malformed names/pseudo-fields, configurable resource budgets,
+header assembly deadlines and idle-stream cancellation. HTTP/1 request/header/body security tests
+remain in the full suite and the shared public message model uses the same bounds on both paths.

flash/docs/http2/STREAMS.md

@@ -0,0 +1,48 @@
+# HTTP/2 streams and request dispatch
+
+Each connection owns a fixed-capacity `Http2StreamTable`. Client stream identifiers are validated
+as odd, non-zero and strictly increasing before a stream object is acquired. The table uses
+primitive open addressing and a bounded object free list; it never grows beyond the advertised 64
+concurrent streams. An excess request receives `REFUSED_STREAM`, allowing the peer to retry it.
+
+## State model
+
+`Http2StreamState` represents `IDLE`, `OPEN`, `HALF_CLOSED_REMOTE`, `HALF_CLOSED_LOCAL` and
+`CLOSED`. A class-initialized table maps every receive/send event to either its next state or the
+correct stream error. Frames racing with a recently closed stream follow RFC 9113 §5.1 rather than
+being rejected uniformly.
+
+## Header and request model
+
+Decoded HPACK fields are copied into storage owned by the stream. Before dispatch,
+`PseudoHeaders` enforces ordering, uniqueness, required request pseudo-fields, lowercase regular
+names, connection-specific-field rejection, the `te: trailers` exception and host/authority
+consistency. Pseudo-fields are not exposed as regular headers; `:authority` is also visible as
+`host` so existing middleware sees the same authority through HTTP/1.1 and HTTP/2.
+
+The stream assembles the existing protocol-neutral `Request`, `RequestLine`, `RequestBody` and
+`HeaderView` models. Path/query splitting, routing, middleware, not-found handling and exception
+handling therefore use the same code as HTTP/1.1. `FastPathRouterImpl` is unchanged.
+
+## Dispatch and ownership
+
+The connection thread decodes and validates frames only. Completed bodyless streams are queued in
+a fixed array while more frame bytes are already buffered, then submitted to the server's shared
+virtual-thread executor before the demultiplexer waits for the network again. This preserves burst
+admission semantics without adding a dispatch timer or blocking the connection thread.
+
+The stream owns its pooled request, response, body, decoded-header arena and response writer.
+Normal response completion releases it through the serialized writer callback. RST_STREAM marks a
+queued or running stream cancelled and defers release to that sole owner; setup, routing and handler
+failures send an appropriate stream reset and release in the failure path. A 100,000-cycle test
+proves stable pool counts, and an immediate request/reset/request regression test covers reuse
+while dispatch is pending.
+
+## Verification
+
+- The complete clean Maven/JMH suite is recorded in `COMPLIANCE.md` and `PERFORMANCE.md`.
+- h2spec sections 5 and 8 pass after request DATA byte accounting landed.
+- Java `HttpClient` negotiates HTTP/2 over TLS and runs an existing parameterized route unchanged.
+- curl prior-knowledge h2c receives a valid `200` response and body.
+- The pooled lifecycle (HPACK decode, request assembly, response write and release) remains a
+  zero-GC CI gate; current percentile and allocation baselines live in `BASELINES.md`.

flash/docs/http2/TROUBLESHOOTING.md

@@ -0,0 +1,73 @@
+# HTTP/2 troubleshooting
+
+## Confirm which protocol was selected
+
+For TLS, the client and server must both offer `h2` through ALPN. Enable
+`FlashConfiguration.http2Enabled`, use a certificate valid for the requested hostname, then check
+with `curl --http2 -v https://host/path` or `nghttp -nv https://host/path`. The trace must report
+ALPN `h2`; a successful HTTP/1.1 response usually means HTTP/2 was not enabled or the client did
+not offer it.
+
+For plaintext, enable `http2CleartextEnabled` and use prior knowledge:
+
+```bash
+curl --http2-prior-knowledge -v http://host:port/path
+nghttp -nv http://host:port/path
+```
+
+Flash does not support `Upgrade: h2c`. A client configured for Upgrade rather than prior knowledge
+will remain on HTTP/1.1.
+
+## Read GOAWAY and RST_STREAM
+
+GOAWAY terminates or drains a connection; `last_stream_id` identifies the highest client stream
+the server may have processed. A client may retry a stream above that id only when its own request
+semantics make retry safe. RST_STREAM affects one stream and leaves the connection usable.
+
+| Error | What it usually means | What to check |
+|---|---|---|
+| `NO_ERROR` | Graceful shutdown or connection rotation. | Server lifecycle and configured connection lifetime. |
+| `PROTOCOL_ERROR` | Invalid preface, pseudo-header ordering, stream state or frame semantics. | A verbose frame trace and the first rejected stream. |
+| `INTERNAL_ERROR` | Handler, response production or I/O failed unexpectedly. | The server exception immediately preceding stream cancellation. |
+| `FLOW_CONTROL_ERROR` | A window overflow or DATA exceeded available credit. | Client flow-control implementation and SETTINGS deltas. |
+| `SETTINGS_TIMEOUT` | The peer did not complete required SETTINGS progress. | Network stalls or a non-compliant peer. |
+| `STREAM_CLOSED` | A frame targeted a stream whose remote side or whole lifecycle was closed. | Late DATA/HEADERS and duplicate terminal frames. |
+| `FRAME_SIZE_ERROR` | A frame length violated its type or the negotiated maximum. | The nine-byte frame header and peer frame-size configuration. |
+| `REFUSED_STREAM` | Live or pending-output capacity was temporarily exhausted. | Client concurrency versus the advertised maximum; retry only when safe. |
+| `CANCEL` | The request, handler or streamed response was cancelled. | Client cancellation and application producer logs. |
+| `COMPRESSION_ERROR` | HPACK integer, Huffman, index or table update was invalid. | Header-block bytes and whether an intermediary rewrote them. |
+| `CONNECT_ERROR` | A CONNECT tunnel failed. | Upstream tunnel or extended-CONNECT negotiation. |
+| `ENHANCE_YOUR_CALM` | A configured abuse, rate, header, body or queue bound was exceeded. | [Security controls](SECURITY.md) and traffic rate before increasing a limit. |
+| `INADEQUATE_SECURITY` | TLS does not meet HTTP/2 requirements. | TLS version, cipher suite and ALPN configuration. |
+| `HTTP_1_1_REQUIRED` | The peer should retry using HTTP/1.1. | Protocol policy and intermediary compatibility. |
+
+Flash caps GOAWAY debug data, and clients must not depend on it being present. The numeric error
+code and last stream id are the reliable diagnostic fields.
+
+## Capture a frame trace
+
+Flash does not log every frame in production: frame logs leak header and traffic metadata and add
+work to the hottest connection loop. Reproduce against a verbose client instead:
+
+```bash
+nghttp -nv https://host/path
+curl --http2 -v https://host/path
+```
+
+`nghttp -nv` prints SETTINGS, HEADERS, DATA, WINDOW_UPDATE, RST_STREAM and GOAWAY in wire order. For
+a server-side-only failure, capture the connection with an approved packet tool; TLS traffic must
+be decrypted in a controlled environment. Never attach production header blocks or payloads to a
+ticket without redacting credentials and personal data.
+
+## Common misconfiguration patterns
+
+1. **HTTP/2 switch disabled.** `http2Enabled` controls TLS ALPN and
+   `http2CleartextEnabled` controls prior knowledge independently.
+2. **Wrong cleartext mode.** The client sends `Upgrade: h2c`; Flash expects the RFC 9113 prior-
+   knowledge preface on the shared plaintext listener.
+3. **ALPN or certificate mismatch.** A custom `TlsConfig.ofContext` omits `h2`, or hostname
+   verification rejects the certificate before HTTP/2 starts. Inspect the TLS handshake first.
+
+If a connection closes under load rather than at startup, compare the observed rate and retained
+stream count with [the security defaults](SECURITY.md), especially reset/stream creation budgets,
+the 64 concurrent-stream setting, header assembly time and stream idle time.

flash/docs/http2/WEBSOCKET.md

@@ -0,0 +1,52 @@
+# WebSockets over HTTP/2
+
+Flash implements RFC 8441 extended CONNECT alongside the existing HTTP/1.1 WebSocket upgrade.
+Both transports resolve the same `ws(path, handler)` registration through `AbstractWsRouter` and
+run the same `WebSocketSession`, frame parser, handler callbacks, and close lifecycle.
+
+## Protocol negotiation
+
+Every HTTP/2 server connection advertises `SETTINGS_ENABLE_CONNECT_PROTOCOL` (`0x8`) with value
+`1`. A WebSocket request uses this pseudo-header shape:
+
+```text
+:method    CONNECT
+:protocol  websocket
+:scheme    https       # or http
+:authority example.com
+:path      /live
+```
+
+The normal HTTP/1.1 upgrade fields (`Connection`, `Upgrade`, `Sec-WebSocket-Key`, and
+`Sec-WebSocket-Accept`) are neither required nor permitted on this path. A matched route receives
+status `200`; a missing route receives `404`.
+
+## Shared application behavior
+
+At the router boundary, an extended CONNECT for `websocket` is represented as a GET so the
+existing WebSocket router can be reused without a second registration table or protocol-specific
+handler API. The wire validator retains the original CONNECT semantics and rejects malformed
+pseudo-header combinations before dispatch.
+
+Request DATA is exposed through the existing streaming `RequestBody`. WebSocket output passes
+through the common push-style `ResponseStream`, so HTTP/2 stream and connection flow-control
+windows apply without changing the WebSocket codec. Messages may cross any number of DATA-frame
+boundaries; those boundaries are invisible to RFC 6455 framing. Client-to-server masking remains
+mandatory and is validated by the same frame parser used for HTTP/1.1.
+
+## Lifecycle and backpressure
+
+Response HEADERS are sent before the push producer is allowed to wait for request DATA. This is
+required for a full-duplex protocol: waiting for the first WebSocket frame before publishing the
+successful CONNECT response would deadlock compliant clients. Subsequent response batches block
+behind the bounded response bridge and resume when HTTP/2 flow-control credit becomes available.
+
+Handler failures from `onOpen` or `onMessage` are reported through `onError`; `onClose` is invoked
+once and the transport is released even if the close callback itself fails.
+
+## Verification
+
+`WebSocketOverH2Test` exercises the extended CONNECT exchange, fragmented text, masking, graceful
+close, and a binary message larger than the initial one-mebibyte stream window.
+`WebSocketParityTest` sends the same message through one route and handler over HTTP/1.1 and
+HTTP/2 and compares the result byte for byte.

flash/docs/http2/WRITER.md

@@ -0,0 +1,275 @@
+# The serialized frame writer
+
+Audience: contributors. This is the design record and benchmark evidence for
+`dev.relism.flash.http2.frame.Http2FrameWriter`, the one component every HTTP/2 write in this
+codebase passes through. It was the central architectural risk: frames, HPACK and flow control are
+table-driven, but multiplexed streams require concurrent producers to share one socket without
+interleaving bytes or pinning carrier threads. The measured gate is recorded below.
+
+## The problem, precisely
+
+Under HTTP/1.1, one virtual thread owns one connection's socket for the request/response it is
+currently serving; there is never a second writer. Under HTTP/2, N streams share one connection
+and their frames must interleave on the wire, so every write must pass through a serialization
+point that plain HTTP/1.1 never needed. A lock taken naively per frame — `synchronized` or an
+uncontended `ReentrantLock.lock()` — costs more per write than every allocation this codebase has
+ever saved elsewhere (`EX-04` through `EX-29`), because it sits on the one path every response,
+of either protocol width, eventually goes through.
+
+## The design, three layers
+
+**Layer 1 — serialize outside the lock.** By the time `Http2FrameWriter.write(WriteIntent)` is
+called, the caller (a stream, or a connection-level singleton such as a precompiled SETTINGS ACK)
+has already built its complete frame — header, HPACK block, payload — into a buffer it owns. The
+writer never serializes anything; it holds the lock only for the duration of one bulk
+`sink.write(buffer, offset, length)` call, never for a sequence of small writes. This is why
+`EX-27` (collapsing `HttpServer.writeResponse`'s ~10 small writes into one) is a prerequisite for
+HTTP/1.1 too; `Http1ResponseWriter` now follows the same bulk-write discipline.
+
+**Layer 2 — `ReentrantLock`, never `synchronized`.** On Java 21, a virtual thread that blocks
+inside a `synchronized` block pins its carrier platform thread (JEP 491, which removes this,
+only lands in JDK 24+). Blocking on a `ReentrantLock` unmounts the virtual thread instead. This
+is the same fix `EX-01` applies to `WebSocketSession`, generalized to the connection writer where
+it matters far more (N streams instead of one WebSocket session). `ReentrantLock` is load-bearing
+for a second reason `synchronized` cannot offer: `tryLock()`.
+
+**Layer 3 — `tryLock()` fast path, intrusive MPSC fallback.** The overwhelmingly common instant,
+even on a genuinely multiplexed connection, has exactly one stream wanting to write: a browser
+calling one API endpoint, a gRPC unary call. `tryLock()` on an uncontended lock is one successful
+CAS; the calling thread writes inline and releases — no handoff, no queue touched, no allocation,
+no context switch. Only when `tryLock()` fails — genuine contention, genuine multiplexing — does
+the intent get published through `IntrusiveMpscQueue` (one more CAS, still zero allocation: the
+`WriteIntent` itself is the queue node, via `mpscNext()`/`setMpscNext`) for the current lock
+holder to drain.
+
+```
+happy path (1 active writer):  tryLock → sink.write → unlock              ≈ 1 CAS
+contended  (N active writers): tryLock fails → CAS enqueue → return
+                               current holder drains the queue before unlocking
+```
+
+### Why the fast path checks `queue.hasWork()`, not just `tryLock()`
+
+Found by this phase's own stress test at N=64/256 — exactly the class of bug R10 exists to catch
+before it ships, not after. Writing an intent immediately, ahead of anything already queued, is
+only safe when nothing is already queued. Without the `hasWork()` guard:
+
+1. Producer P calls `write(a)`, then `write(b)`. Both contend (someone else holds the lock) and
+   both get queued — fire-and-forget from P's point of view.
+2. The current holder is *about* to drain them but has not yet done so.
+3. P's very next call, `write(c)`, finds the lock free (the holder released it between P's calls)
+   and — without the guard — would write `c` directly, landing it on the wire *before* `a` and
+   `b`, which are still sitting in the queue.
+
+`write()` therefore checks `!queue.hasWork() && lock.tryLock()` before taking the direct path:
+"bypass the queue" only happens when the queue is observed genuinely empty, i.e. everything any
+producer has ever offered has already been written. `hasWork()` never false-negatives (it would
+only ever wrongly report work that isn't there, which just costs an extra harmless `tryLock()`
+attempt), so this preserves per-producer ordering without adding a false rejection of the fast
+path.
+
+## Lost-wakeup avoidance
+
+The classic hazard for a design like this: a producer offers its intent to the queue at the exact
+moment the current lock holder has just found the queue empty and is about to unlock. Without
+care, the item is stranded — offered, but nobody left to drain it, and the producer already
+returned believing the write is in flight.
+
+```
+Producer P                              Holder H (currently draining, about to unlock)
+───────────                             ──────────────────────────────────────────────
+                                         next = queue.poll()   // null: queue looks empty
+queue.offer(intent)     ← races here →
+if (lock.tryLock())                     lock.unlock()
+    drive(null)          // P's own second chance: if P wins the tryLock() race
+                          // immediately after H's unlock(), P itself becomes the new
+                          // holder and drains — including its own just-offered intent.
+```
+
+Two cooperating mechanisms close this, and both are required — neither alone is sufficient:
+
+1. **The producer's own second chance.** After a failed `tryLock()`, `write()` offers the intent
+   *then* immediately attempts `tryLock()` again. If H has already unlocked by this point, P wins
+   the second `tryLock()` and drains the queue itself (`drive(null)` — draining whatever is
+   queued, which necessarily includes the intent P just offered, since `offer()` had
+   already completed).
+2. **The holder's re-check-after-unlock loop**, in `drive()`: after `unlock()`, re-read
+   `queue.hasWork()`. If non-empty, attempt `tryLock()` again and drain, then unlock and re-check
+   once more — looping, because this recheck cycle can itself race the same way a first pass can.
+   If a second `tryLock()` in this loop fails, some *other* thread now holds the lock, and by the
+   same argument that other holder's own re-check-after-unlock covers the item once it releases.
+
+The correctness argument for why together these are sufficient is a happens-before chain through
+the queue's `AtomicReference` (`IntrusiveMpscQueue.head`, a `getAndSet` per `offer`) and the
+lock's own acquire/release ordering: every `offer()` happens-before some subsequent `poll()` that
+observes it (directly, or via the momentary-`null` self-correcting race documented on
+`IntrusiveMpscQueue` itself — see its class Javadoc), and every thread that successfully offers
+either (a) is itself about to attempt `tryLock()` and, on success, drains everything including its
+own offer, or (b) fails that `tryLock()`, meaning some other thread holds the lock *at that
+instant* and that thread's own unlock will trigger its own re-check-after-unlock loop. There is no
+interleaving in which an offered intent is neither drained by its own producer nor covered by some
+other thread's re-check loop.
+
+A frame's bytes are never interleaved with another frame's bytes: every write of one intent is a
+single `sink.write` call issued while holding the lock, and the lock is not released between a
+`WriteIntent`'s bytes — proven directly by `Http2FrameWriterStressTest`, which reassembles
+producer/sequence/marker-tagged frames from the sink's output and fails loudly on any torn,
+duplicated, reordered, or lost frame.
+
+## Write timeout
+
+A blocking write is unavoidable when the kernel send buffer is full and the peer is not reading —
+whoever holds the lock is blocked in the syscall, holding up every other stream on the connection.
+This is bounded by `Http2Limits.WRITE_TIMEOUT_MS` (30 s), enforced by a single shared daemon
+thread (`Http2FrameWriter.WriteTimeoutReaper`) rather than `Socket#setSoTimeout`, which bounds
+reads, not writes.
+
+The reaper deliberately does **not** ask each write to record a `System.nanoTime()` deadline — an
+early revision did, and this phase's own N=1 benchmark measured that single `nanoTime()` call
+(plus the extra `volatile` field it required) costing enough to put per-write overhead over the
+50 ns-over-baseline gate budget. Instead, the reaper scans every registered writer every
+`SCAN_INTERVAL_MS` (50 ms) and counts *consecutive* scans a writer has been observed still blocked
+(`writingThread` non-null); a writer blocked for more than `WRITE_TIMEOUT_MS / SCAN_INTERVAL_MS`
+consecutive scans is interrupted. This trades a little precision — up to one scan interval of
+slop, already inherent to any background-reaper design — for removing all per-write timing cost
+from the path this document's gate criteria are strictest about.
+
+## Benchmark methodology
+
+`flash/src/jmh/java/dev/relism/flash/http2/frame/FrameWriterBenchmark.java` (a JMH source root
+registered only under the `jmh` Maven profile, not `src/test/java`) compares four harnesses at
+`threads` ∈ {1, 2, 4, 8, 16, 64}:
+
+- `trylock_mpsc` — the shipped `Http2FrameWriter` design.
+- `plain_lock` — every write blocks on `ReentrantLock.lock()` unconditionally (candidate (a)).
+- `dedicated_thread` — every write hands off to one dedicated platform thread via the same
+  `IntrusiveMpscQueue`, parked/unparked, never busy-polled (candidate (c)).
+- `raw_unsynchronized` — no coordination at all; not a candidate (concurrent writers would tear
+  each other's frames), included only to answer "what does a write cost with zero coordination",
+  which the N=1 gate criterion is defined relative to.
+
+Each JMH "operation" is a full burst: `threads` virtual producer threads each write 4 000 frames
+of 512 bytes into a `CountingSink` that discards the bytes but atomically counts completed writes;
+`runBurst` blocks until the count reaches the expected total, so the timed interval always covers
+real completion, not mere submission (`write()` can return once an intent is merely *queued* on
+the contended path — timing only "how long until every `write()` call returned" would flatter
+whichever design most aggressively defers work). `@Threads` was not usable here: it requires a
+compile-time constant, not a value swept via `@Param`, and JMH's own thread pool is platform
+threads, not the virtual threads under test.
+
+**Two honest caveats, stated plainly rather than glossed over (R3):**
+
+1. **The sink is an in-memory counter, not a real socket.** "p999 latency ... on loopback" in the
+   plan's gate wording implies real socket I/O; this harness measures writer-lock-contention
+   latency in isolation from network variance, which is the right isolation for judging *this
+   component*, but it means the recorded p999 numbers below are a lower bound on what a real
+   loopback socket would show, not a direct stand-in for it. Frame size used is 512 B, not the
+   plan's illustrative 1 KB — chosen to keep the burst's own array allocation small relative to
+   JVM defaults; the writer's cost model does not depend on frame size (it copies nothing; see
+   `WriteIntent`'s Javadoc), so this does not affect the gate conclusions.
+2. **One JMH "op" is a whole burst (4 000 writes), not one write**, because `@OperationsPerInvocation`
+   requires a compile-time constant and cannot vary with the `threads` `@Param`. Every burst also
+   pays fixed harness costs common to *all four* designs equally: one `ExecutorService` (a
+   virtual-thread-per-task executor) created and torn down, one `Future[]` array, one
+   `long[threads][4000]` latency-sample array, and one fresh `BenchIntent` object allocated per
+   write (matching the stress test's own pattern, not the writer's actual production contract —
+   a real stream is long-lived and reuses itself as its own `WriteIntent`). Because this cost is
+   identical across designs, **absolute** `gc.alloc.rate.norm` numbers below are dominated by this
+   shared harness cost (~33 443 B/op), not by the design under test; the number that actually
+   answers the "0 B/op" gate criterion is the **differential** between a design and the
+   `raw_unsynchronized` baseline, which isolates exactly the bytes that design itself adds.
+
+## Results
+
+All runs: JDK 21.0.11 (Temurin), this development sandbox, JMH 1.37, `-Fork` per run noted below.
+Raw JMH output is not reproduced in full here; the numbers below are the reported means with
+their 99.9% CI half-widths.
+
+### N=1 — throughput and allocation (`-f 4 -wi 5 -w 1s -i 12 -r 2s`, throughput; separately
+`-f 2 -wi 3 -i 8`, `-prof gc`)
+
+| design | ops/s (bursts/s) | derived ns/write | gc.alloc.rate.norm (B/op, per burst) |
+|---|---|---|---|
+| `trylock_mpsc` | 2007.930 ± 60.623 | 124.5 ns | 33 449.253 ± 13.383 |
+| `raw_unsynchronized` | 3052.036 ± 52.515 | 81.9 ns | 33 443.349 ± 1.572 |
+
+- **Overhead vs. raw unsynchronized:** 124.5 − 81.9 = **42.6 ns** (point estimate). Worst case
+  within the 99.9% CI (slowest plausible `trylock_mpsc`, fastest plausible baseline):
+  ≈ **47.9 ns**. Both are under the **50 ns** gate budget.
+- **Allocation delta:** 33 449.253 − 33 443.349 = **5.9 B per 4 000-write burst** ≈ **0.0015 B per
+  write** — within `trylock_mpsc`'s own ±13.383 error band, i.e. not distinguishable from zero.
+  Consistent with the design: the fast path is `queue.hasWork()` (a volatile read) plus
+  `ReentrantLock.tryLock()`/`unlock()` (well-known non-allocating on the JDK's implementation)
+  plus one bulk `sink.write`. **Gate criterion: 0 B/op — PASS.**
+
+### N=64 — throughput retention and tail latency (`-f 2 -wi 3 -w 1s -i 5 -r 1s`)
+
+| design | N=1 writes/s (per-thread) | N=64 writes/s (aggregate) | retention | p999 @ N=64 |
+|---|---|---|---|---|
+| `trylock_mpsc` (shipped) | 8 721 148 | 5 712 640 | **65.5 %** | **11.8–14.2 µs** |
+| `plain_lock` (candidate a) | 10 469 956 | 6 082 048 | 58.1 % | 1627–1952 µs |
+| `dedicated_thread` (candidate c) | 2 673 964 | 5 718 528 | 213.8 %† | 1.5–6.7 µs |
+| `raw_unsynchronized` (unsafe baseline) | 11 353 924 | 20 764 160 | n/a | n/a |
+
+† `dedicated_thread`'s N=1 baseline is itself poor (every uncontended write still pays a full
+park/unpark handoff to the dedicated thread — there is no fast path for the "only one writer"
+case at all), so a >100% "retention" number reflects a bad denominator, not superlinear scaling.
+It is reported for completeness, not as a pass/fail signal — the gate criterion is defined
+relative to `trylock_mpsc`'s own N=1 baseline, which is the design that shipped.
+
+- **`trylock_mpsc` throughput retention:** 65.5 % ≥ the required 60 %. **PASS.**
+- **`trylock_mpsc` p999 latency:** 11.8–14.2 µs, far under the 1 ms budget. **PASS.**
+- (Not gate-relevant, but part of why (b) was chosen over (a) and (c), per the plan's task 6:
+  `plain_lock` blows past the 1 ms p999 budget by ~1000× under load — unfair blocking causes tail
+  pile-up exactly as expected from a design with no fast path and no fairness guarantee.
+  `dedicated_thread` has the best tail latency of the three but a **~3.3×** throughput penalty at
+  N=1, because *every* write, even genuinely uncontended ones, pays a full thread handoff. Neither
+  alternative is a better shipped default than `trylock_mpsc`.)
+
+### Stress test — correctness under concurrency, 1000 iterations per N
+
+Run via an ad hoc reflective driver invoking `Http2FrameWriterStressTest`'s private `runStress`
+method directly (the shipped test class runs reduced counts for a fast default `mvn test`; this
+is the full gate verification described in that class's own Javadoc), for `N` ∈ {1, 2, 8, 64,
+256}, 1000 iterations each:
+
+| Scheduler | n=1 | n=2 | n=8 | n=64 | n=256 | Total wall time |
+|---|---|---|---|---|---|---|
+| default parallelism | 0 failures | 0 failures | 0 failures | 0 failures | 0 failures | ≈ 9.1 s |
+| `-Djdk.virtualThreadScheduler.parallelism=1` | 0 failures | 0 failures | 0 failures | 0 failures | 0 failures | ≈ 8.9 s |
+
+Every byte of every frame arrived, correctly ordered per-producer, with no tearing, duplication,
+or loss, in both configurations — **10 000 total stress runs, 0 failures.**
+
+**Carrier pinning:** the `parallelism=1` run above was additionally run under
+`-Djdk.tracePinnedThreads=full`, which prints a stack trace to stderr for any virtual thread found
+blocked while pinning its carrier. Zero output — **no pinning observed**, consistent with the
+design's exclusive use of `ReentrantLock` (never `synchronized`) on every path that can block.
+
+## Gate criteria — final tally
+
+| # | Criterion | Result | Verdict |
+|---|---|---|---|
+| 1 | N=1: 0 B/op | 0.0015 B/write differential vs. baseline, within noise | **PASS** |
+| 1 | N=1: ≤50 ns overhead vs. raw unsynchronized | 42.6 ns point estimate, ≤47.9 ns worst-case CI | **PASS** |
+| 2 | N=64: throughput ≥60% of N=1 per-thread rate | 65.5 % | **PASS** |
+| 2 | N=64: p999 <1 ms (512 B frame, in-memory sink) | 11.8–14.2 µs | **PASS** |
+| 3 | No carrier pinning under `-Djdk.tracePinnedThreads=full` | none observed | **PASS** |
+| 4 | Stress test green at every N, 1000 iterations, incl. parallelism=1 | 10 000/10 000 | **PASS** |
+
+**All four gate criteria are met.** `Http2FrameWriter` ships as designed: a `tryLock()` fast path
+with an intrusive MPSC fallback.
+
+## What this design costs vs. what it saves
+
+The honest framing (per R3, extended from HPACK's own to the writer): the writer's happy path
+costs one uncontended CAS (`ReentrantLock.tryLock()`) plus a volatile read (`queue.hasWork()`)
+plus the write syscall itself — on the order of tens of nanoseconds, measured above at ~42.6 ns
+over a raw unsynchronized write. What it buys is the only thing that makes HTTP/2 multiplexing
+possible on a codebase built around "one thread owns the socket": N concurrent streams can write
+frames to the same connection without a naive per-frame lock (which the `plain_lock` comparison
+above shows costs ~1000× more in tail latency once real contention appears), and without
+committing every connection to a dedicated writer thread's per-write handoff cost (which the
+`dedicated_thread` comparison shows costs ~3.3× throughput at the N=1 case that dominates real
+traffic). Forty-two nanoseconds is a price worth paying once, on the one path that gates
+multiplexed HTTP/2 correctness at all.

flash/pom.xml

@@ -37,4 +37,82 @@
         </dependency>
     </dependencies>
 
+    <!--
+      Phase 3 (flash/docs/http2/IMPLEMENTATION-PLAN.md): the JMH benchmark gate for the h2
+      serialized frame writer. Not bound to the default build — activate explicitly with
+      `-Pjmh`. Benchmarks live in src/jmh/java, a source root distinct from src/test/java
+      (a `jmh` profile on this module, per the plan's own suggestion, rather than a new
+      flash-bench submodule — recorded as DEC-09), specifically so that `mvn test` with no
+      profile never even *compiles* them: src/jmh/java is registered as a test-source root
+      only inside this profile (build-helper-maven-plugin's add-test-source), and the JMH
+      dependencies it imports are likewise profile-scoped. An earlier revision put the
+      benchmark directly in src/test/java, relying on Surefire's JUnit filtering (JMH classes
+      carry no JUnit annotations) to skip it at *run* time — but Surefire's test discovery
+      loads every compiled test class regardless, so a plain `mvn test` without `-Pjmh` failed
+      the whole module at test-compile with "package org.openjdk.jmh.annotations does not
+      exist", since jmh-core is not on the classpath outside this profile. The separate source
+      root fixes that at the root: with the profile inactive, the benchmark source is not on
+      any compiler's input at all. Run: `mvn -Pjmh -pl flash test-compile` then
+      `java -cp ... org.openjdk.jmh.Main` (see FrameWriterBenchmark's own Javadoc for the full
+      classpath incantation).
+    -->
+    <profiles>
+        <profile>
+            <id>jmh</id>
+            <dependencies>
+                <dependency>
+                    <groupId>org.openjdk.jmh</groupId>
+                    <artifactId>jmh-core</artifactId>
+                    <version>${jmh.version}</version>
+                </dependency>
+                <dependency>
+                    <groupId>org.openjdk.jmh</groupId>
+                    <artifactId>jmh-generator-annprocess</artifactId>
+                    <version>${jmh.version}</version>
+                </dependency>
+            </dependencies>
+            <build>
+                <plugins>
+                    <plugin>
+                        <groupId>org.codehaus.mojo</groupId>
+                        <artifactId>build-helper-maven-plugin</artifactId>
+                        <version>${build.helper.plugin.version}</version>
+                        <executions>
+                            <execution>
+                                <id>add-jmh-source</id>
+                                <phase>generate-test-sources</phase>
+                                <goals>
+                                    <goal>add-test-source</goal>
+                                </goals>
+                                <configuration>
+                                    <sources>
+                                        <source>src/jmh/java</source>
+                                    </sources>
+                                </configuration>
+                            </execution>
+                        </executions>
+                    </plugin>
+                    <plugin>
+                        <groupId>org.apache.maven.plugins</groupId>
+                        <artifactId>maven-compiler-plugin</artifactId>
+                        <configuration>
+                            <annotationProcessorPaths>
+                                <path>
+                                    <groupId>org.projectlombok</groupId>
+                                    <artifactId>lombok</artifactId>
+                                    <version>${lombok.version}</version>
+                                </path>
+                                <path>
+                                    <groupId>org.openjdk.jmh</groupId>
+                                    <artifactId>jmh-generator-annprocess</artifactId>
+                                    <version>${jmh.version}</version>
+                                </path>
+                            </annotationProcessorPaths>
+                        </configuration>
+                    </plugin>
+                </plugins>
+            </build>
+        </profile>
+    </profiles>
+
 </project>

flash/src/bench/java/dev/relism/flash/bench/BenchmarkMain.java

@@ -0,0 +1,69 @@
+package dev.relism.flash.bench;
+
+import dev.relism.flash.extension.FlashApp;
+import dev.relism.flash.extension.FlashConfiguration;
+import java.net.ServerSocket;
+import java.net.URI;
+import java.time.Duration;
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * Real-server, real-network throughput and latency benchmark: boots one live Flash server on
+ * loopback exposing {@code GET /hello}, then drives it end to end — real sockets, real accept
+ * loop, real routing and response serialization — with independent HTTP clients across protocols
+ * and concurrency levels. This is not a component-scoped JMH microbenchmark; it is the same shape
+ * of measurement a tool like {@code h2load} or {@code wrk} gives any other server.
+ *
+ * <p>Never wired into the build or CI — run manually with: {@code mvn -pl flash -Pbench
+ * exec:java}. Override scenario length with {@code -Dflash.bench.warmupSeconds} / {@code
+ * -Dflash.bench.measureSeconds} (defaults: 2 / 5).
+ */
+public final class BenchmarkMain {
+
+  private static final int[] CONCURRENCY_LEVELS = {1, 8, 32, 128};
+
+  public static void main(String[] args) throws Exception {
+    Duration warmup = seconds("flash.bench.warmupSeconds", 2);
+    Duration measurement = seconds("flash.bench.measureSeconds", 5);
+
+    int port = freePort();
+    FlashApp app =
+        FlashApp.create(
+            FlashConfiguration.builder()
+                .host("127.0.0.1")
+                .port(port)
+                .http2CleartextEnabled(true)
+                .build());
+    app.get("/hello", (request, response) -> "hello");
+    app.start();
+
+    try {
+      URI target = URI.create("http://127.0.0.1:" + port + "/hello");
+      Report.print(runAllScenarios(target, warmup, measurement));
+    } finally {
+      app.stop().join();
+    }
+  }
+
+  private static List<LoadResult> runAllScenarios(URI target, Duration warmup, Duration measurement)
+      throws InterruptedException {
+    List<LoadResult> results = new ArrayList<>();
+    for (int concurrency : CONCURRENCY_LEVELS) {
+      results.add(
+          new Http1Driver()
+              .run("http/1.1 c=" + concurrency, target, concurrency, warmup, measurement));
+    }
+    return results;
+  }
+
+  private static Duration seconds(String property, int fallback) {
+    return Duration.ofSeconds(Long.getLong(property, fallback));
+  }
+
+  private static int freePort() throws Exception {
+    try (ServerSocket socket = new ServerSocket(0)) {
+      return socket.getLocalPort();
+    }
+  }
+}

flash/src/bench/java/dev/relism/flash/bench/Http1Driver.java

@@ -0,0 +1,37 @@
+package dev.relism.flash.bench;
+
+import java.net.URI;
+import java.net.http.HttpClient;
+import java.net.http.HttpRequest;
+import java.net.http.HttpResponse;
+import java.time.Duration;
+
+/**
+ * HTTP/1.1 keep-alive load driver backed by the JDK's own {@link HttpClient} — an independent
+ * client implementation, not Flash's own code, measuring the server end to end.
+ */
+final class Http1Driver implements LoadDriver {
+
+  @Override
+  public LoadResult run(
+      String scenarioLabel, URI target, int concurrency, Duration warmup, Duration measurement)
+      throws InterruptedException {
+    HttpRequest request = HttpRequest.newBuilder(target).timeout(Duration.ofSeconds(5)).GET().build();
+    return LoadRunner.execute(
+        scenarioLabel,
+        concurrency,
+        warmup,
+        measurement,
+        () -> {
+          // One HttpClient per worker: its own connection pool, reused keep-alive across requests.
+          HttpClient client = HttpClient.newBuilder().version(HttpClient.Version.HTTP_1_1).build();
+          return () -> {
+            HttpResponse<Void> response =
+                client.send(request, HttpResponse.BodyHandlers.discarding());
+            if (response.statusCode() != 200) {
+              throw new IllegalStateException("status " + response.statusCode());
+            }
+          };
+        });
+  }
+}

flash/src/bench/java/dev/relism/flash/bench/LatencyRecorder.java

@@ -0,0 +1,22 @@
+package dev.relism.flash.bench;
+
+import java.util.Arrays;
+
+/** One worker's latency samples, in nanoseconds. Grows without boxing on the request loop. */
+final class LatencyRecorder {
+  private long[] samples = new long[1024];
+  private int count;
+
+  void record(long nanos) {
+    if (count == samples.length) samples = Arrays.copyOf(samples, samples.length * 2);
+    samples[count++] = nanos;
+  }
+
+  int count() {
+    return count;
+  }
+
+  long[] toArray() {
+    return Arrays.copyOf(samples, count);
+  }
+}

flash/src/bench/java/dev/relism/flash/bench/LoadDriver.java

@@ -0,0 +1,11 @@
+package dev.relism.flash.bench;
+
+import java.net.URI;
+import java.time.Duration;
+
+/** Runs one scenario (a protocol at a fixed concurrency) against a live target and returns its stats. */
+interface LoadDriver {
+  LoadResult run(
+      String scenarioLabel, URI target, int concurrency, Duration warmup, Duration measurement)
+      throws InterruptedException;
+}

flash/src/bench/java/dev/relism/flash/bench/LoadResult.java

@@ -0,0 +1,17 @@
+package dev.relism.flash.bench;
+
+/** One scenario's outcome: throughput and latency distribution over the measured phase only. */
+record LoadResult(
+    String scenario,
+    long requests,
+    long errors,
+    double seconds,
+    double meanLatencyMicros,
+    double p50Micros,
+    double p99Micros,
+    double p999Micros) {
+
+  double requestsPerSecond() {
+    return seconds == 0 ? 0 : requests / seconds;
+  }
+}

flash/src/bench/java/dev/relism/flash/bench/LoadRunner.java

@@ -0,0 +1,69 @@
+package dev.relism.flash.bench;
+
+import java.time.Duration;
+import java.util.ArrayList;
+import java.util.List;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+import java.util.concurrent.atomic.LongAdder;
+
+/**
+ * Drives a fixed number of concurrent virtual-thread workers against one {@link WorkerFactory},
+ * each worker looping its own {@link WorkUnit#request()} until a wall-clock deadline. A discarded
+ * warmup phase runs first so JIT warmup and connection setup don't skew the measured phase.
+ */
+final class LoadRunner {
+
+  private LoadRunner() {}
+
+  static LoadResult execute(
+      String scenarioLabel,
+      int concurrency,
+      Duration warmup,
+      Duration measurement,
+      WorkerFactory factory)
+      throws InterruptedException {
+    runUntil(concurrency, System.nanoTime() + warmup.toNanos(), factory, null, null);
+
+    LongAdder errors = new LongAdder();
+    List<LatencyRecorder> perWorker = new ArrayList<>(concurrency);
+    for (int i = 0; i < concurrency; i++) perWorker.add(new LatencyRecorder());
+
+    long measureStart = System.nanoTime();
+    runUntil(concurrency, measureStart + measurement.toNanos(), factory, errors, perWorker);
+
+    return Stats.summarize(scenarioLabel, perWorker, errors.sum(), System.nanoTime() - measureStart);
+  }
+
+  private static void runUntil(
+      int concurrency,
+      long deadlineNanos,
+      WorkerFactory factory,
+      LongAdder errors,
+      List<LatencyRecorder> perWorker)
+      throws InterruptedException {
+    try (ExecutorService pool = Executors.newVirtualThreadPerTaskExecutor()) {
+      for (int i = 0; i < concurrency; i++) {
+        LatencyRecorder recorder = perWorker == null ? null : perWorker.get(i);
+        pool.execute(() -> worker(deadlineNanos, factory, errors, recorder));
+      }
+    }
+  }
+
+  private static void worker(
+      long deadlineNanos, WorkerFactory factory, LongAdder errors, LatencyRecorder recorder) {
+    try (WorkUnit unit = factory.create()) {
+      while (System.nanoTime() < deadlineNanos) {
+        long start = System.nanoTime();
+        try {
+          unit.request();
+          if (recorder != null) recorder.record(System.nanoTime() - start);
+        } catch (Exception requestFailure) {
+          if (errors != null) errors.increment();
+        }
+      }
+    } catch (Exception setupFailure) {
+      if (errors != null) errors.increment();
+    }
+  }
+}

flash/src/bench/java/dev/relism/flash/bench/Report.java

@@ -0,0 +1,27 @@
+package dev.relism.flash.bench;
+
+import java.util.List;
+
+/** Prints results as a fixed-width table on stdout — no file output, this is a manual tool. */
+final class Report {
+
+  private Report() {}
+
+  static void print(List<LoadResult> results) {
+    System.out.printf(
+        "%-16s %10s %8s %12s %10s %10s %10s %10s%n",
+        "scenario", "requests", "errors", "req/s", "mean(us)", "p50(us)", "p99(us)", "p999(us)");
+    for (LoadResult result : results) {
+      System.out.printf(
+          "%-16s %10d %8d %12.1f %10.1f %10.1f %10.1f %10.1f%n",
+          result.scenario(),
+          result.requests(),
+          result.errors(),
+          result.requestsPerSecond(),
+          result.meanLatencyMicros(),
+          result.p50Micros(),
+          result.p99Micros(),
+          result.p999Micros());
+    }
+  }
+}

flash/src/bench/java/dev/relism/flash/bench/Stats.java

@@ -0,0 +1,52 @@
+package dev.relism.flash.bench;
+
+import java.util.Arrays;
+import java.util.List;
+
+/** Merges every worker's samples and reduces them to one {@link LoadResult}. */
+final class Stats {
+
+  private Stats() {}
+
+  static LoadResult summarize(
+      String scenarioLabel, List<LatencyRecorder> perWorker, long errors, long elapsedNanos) {
+    int total = 0;
+    for (LatencyRecorder recorder : perWorker) total += recorder.count();
+
+    long[] merged = new long[total];
+    int offset = 0;
+    for (LatencyRecorder recorder : perWorker) {
+      long[] samples = recorder.toArray();
+      System.arraycopy(samples, 0, merged, offset, samples.length);
+      offset += samples.length;
+    }
+    Arrays.sort(merged);
+
+    return new LoadResult(
+        scenarioLabel,
+        merged.length,
+        errors,
+        elapsedNanos / 1_000_000_000.0,
+        microsOf(mean(merged)),
+        microsOf(percentile(merged, 0.50)),
+        microsOf(percentile(merged, 0.99)),
+        microsOf(percentile(merged, 0.999)));
+  }
+
+  private static double mean(long[] sorted) {
+    if (sorted.length == 0) return 0;
+    long sum = 0;
+    for (long value : sorted) sum += value;
+    return (double) sum / sorted.length;
+  }
+
+  private static long percentile(long[] sorted, double fraction) {
+    if (sorted.length == 0) return 0;
+    int index = (int) Math.min(sorted.length - 1, Math.floor(fraction * sorted.length));
+    return sorted[index];
+  }
+
+  private static double microsOf(double nanos) {
+    return nanos / 1000.0;
+  }
+}

flash/src/bench/java/dev/relism/flash/bench/WorkUnit.java

@@ -0,0 +1,9 @@
+package dev.relism.flash.bench;
+
+/** One worker's request loop body. {@link #close()} releases whatever {@link WorkerFactory} opened. */
+interface WorkUnit extends AutoCloseable {
+  void request() throws Exception;
+
+  @Override
+  default void close() throws Exception {}
+}

flash/src/bench/java/dev/relism/flash/bench/WorkerFactory.java

@@ -0,0 +1,7 @@
+package dev.relism.flash.bench;
+
+/** Builds one worker's {@link WorkUnit} — its own connection/client, isolated per virtual thread. */
+@FunctionalInterface
+interface WorkerFactory {
+  WorkUnit create() throws Exception;
+}

flash/src/jmh/java/dev/relism/flash/RequestPipelineBenchmark.java

@@ -0,0 +1,123 @@
+package dev.relism.flash;
+
+import dev.relism.flash.http.HttpMethod;
+import dev.relism.flash.models.Request;
+import dev.relism.flash.models.RequestHandler;
+import dev.relism.flash.models.SimpleHandler;
+import dev.relism.flash.routing.Middleware;
+import dev.relism.flash.routing.routers.fastpathrouter.FastPathRouterImpl;
+import dev.relism.flash.transport.BufferedByteSource;
+import java.io.IOException;
+import java.io.InputStream;
+import java.nio.charset.StandardCharsets;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/**
+ * The h1 zero-alloc contract: "an h1 {@code GET /users/{id}} request that reads three headers and
+ * one path param must be 0 B/op end to end except for the user-facing {@code String}s the handler
+ * explicitly asks for." This benchmark measures the actual number with {@code -prof gc}. Before
+ * model and view pooling, {@code parseAndRoute} measured 120.008 B/op. It now measures at JMH's
+ * allocation noise floor. The two methods below isolate the parser/router path from the unavoidable
+ * cost of explicit {@code String} reads by comparing a route with no header or parameter access
+ * against one that reads a path parameter and two headers.
+ *
+ * <p>Uses a hand-rolled repeating {@link InputStream} (never allocates, cycles the same request
+ * bytes indefinitely) rather than a fresh {@code ByteArrayInputStream}/{@code BufferedByteSource}
+ * per invocation, so the timed path matches production exactly: one {@link BufferedByteSource}
+ * created once per connection and reused across every request, per {@code Http1Connection}'s own
+ * shape — not recreated per benchmark iteration, which would contaminate the measurement with
+ * harness allocation unrelated to the parser/router/model code under test (the same lesson {@code
+ * WRITER.md} documents for {@code FrameWriterBenchmark}'s own harness).
+ */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class RequestPipelineBenchmark {
+
+  /**
+   * Cycles a fixed byte[] indefinitely — simulates an infinite pipelined keep-alive stream of
+   * identical requests without allocating anything per read.
+   */
+  private static final class RepeatingByteStream extends InputStream {
+    private final byte[] template;
+    private int pos;
+
+    RepeatingByteStream(byte[] template) {
+      this.template = template;
+    }
+
+    @Override
+    public int read() {
+      byte b = template[pos];
+      pos = (pos + 1) % template.length;
+      return b & 0xFF;
+    }
+
+    @Override
+    public int read(byte[] dst, int off, int len) {
+      for (int i = 0; i < len; i++) {
+        dst[off + i] = template[pos];
+        pos = (pos + 1) % template.length;
+      }
+      return len;
+    }
+  }
+
+  private RequestParser parser;
+  private BufferedByteSource in;
+  private FastPathRouterImpl router;
+  private Object routeScratch;
+
+  @Setup(Level.Trial)
+  public void setup() {
+    String req =
+        "GET /users/12345 HTTP/1.1\r\n"
+            + "Host: api.example.com\r\n"
+            + "Accept: application/json\r\n"
+            + "Authorization: Bearer abcdefghijklmnopqrstuvwxyz\r\n"
+            + "\r\n";
+    byte[] template = req.getBytes(StandardCharsets.US_ASCII);
+    in = new BufferedByteSource(new RepeatingByteStream(template), null);
+    parser = new RequestParser(64 * 1024);
+
+    router = new FastPathRouterImpl();
+    RequestHandler handler = new SimpleHandler((r, res) -> "ok");
+    router.doRegister(HttpMethod.GET, "/users/{id}", handler, new Middleware[0]);
+    router.compile();
+    routeScratch = router.newScratch();
+  }
+
+  /** Parse and route without requesting user-facing header or parameter strings. */
+  @Benchmark
+  public RequestHandler parseAndRoute() throws IOException {
+    Request request = parser.parse(in);
+    request.drain();
+    return router.route(request, routeScratch);
+  }
+
+  /** Parse, route, and read one path parameter and two headers as strings. */
+  @Benchmark
+  public Object parseRouteAndExtractThreeFields() throws IOException {
+    Request request = parser.parse(in);
+    RequestHandler handler = router.route(request, routeScratch);
+    String id = request.param("id");
+    String host = request.header("Host");
+    String auth = request.header("Authorization");
+    request.drain();
+    return id.length() + host.length() + auth.length() + (handler != null ? 1 : 0);
+  }
+}

flash/src/jmh/java/dev/relism/flash/bytes/ByteScanBenchmark.java

@@ -0,0 +1,62 @@
+package dev.relism.flash.bytes;
+
+import java.nio.charset.StandardCharsets;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/**
+ * Compares {@link ByteScan#indexOfCrLfCrLf} (SWAR) against {@link ByteScan#indexOfCrLfCrLfScalar}
+ * on a realistic HTTP/1.1 request header block. It lives in this package to reach the
+ * package-private scalar reference method without widening that method's visibility solely for
+ * measurement.
+ *
+ * <p>Run: {@code mvn -Pjmh -pl flash test-compile} then {@code java -cp
+ * flash/target/test-classes:flash/target/classes:$(mvn -Pjmh -pl flash dependency:build-classpath
+ * -Dmdep.outputFile=/dev/stdout -q) org.openjdk.jmh.Main ByteScanBenchmark}.
+ */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class ByteScanBenchmark {
+
+  /** A realistic request: request line + 7 headers + terminator, ~330 bytes. */
+  private byte[] requestBuf;
+
+  @Setup(Level.Trial)
+  public void setup() {
+    String req =
+        "GET /users/12345?sort=desc&limit=20 HTTP/1.1\r\n"
+            + "Host: api.example.com\r\n"
+            + "User-Agent: Mozilla/5.0 (compatible; FlashBench/1.0)\r\n"
+            + "Accept: application/json\r\n"
+            + "Accept-Encoding: gzip, deflate, br\r\n"
+            + "Authorization: Bearer abcdefghijklmnopqrstuvwxyz0123456789\r\n"
+            + "Cookie: session=xyz123abc; theme=dark; lang=en-US\r\n"
+            + "Connection: keep-alive\r\n"
+            + "\r\n";
+    requestBuf = req.getBytes(StandardCharsets.US_ASCII);
+  }
+
+  @Benchmark
+  public int headerEndScan_swar() {
+    return ByteScan.indexOfCrLfCrLf(requestBuf, 0, requestBuf.length);
+  }
+
+  @Benchmark
+  public int headerEndScan_scalar() {
+    return ByteScan.indexOfCrLfCrLfScalar(requestBuf, 0, requestBuf.length);
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/Http2ConnectionBenchmark.java

@@ -0,0 +1,131 @@
+package dev.relism.flash.http2;
+
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http2.frame.FrameFlags;
+import dev.relism.flash.http2.frame.FrameType;
+import dev.relism.flash.http2.frame.FrameWriteBuffer;
+import dev.relism.flash.http2.frame.Http2FrameReader;
+import dev.relism.flash.http2.frame.Http2FrameWriter;
+import dev.relism.flash.transport.BufferedByteSource;
+import java.io.InputStream;
+import java.nio.charset.StandardCharsets;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.TearDown;
+import org.openjdk.jmh.annotations.Warmup;
+
+/** Measures the allocation-free control-frame lifecycle after connection objects are prepared. */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class Http2ConnectionBenchmark {
+  private static final java.util.function.BooleanSupplier RUNNING = () -> false;
+
+  private byte[] wire;
+  private Http2Connection connection;
+  private BufferedByteSource input;
+  private Http2FrameReader reader;
+  private Http2FrameWriter writer;
+  private CountingSink sink;
+  private ResettableInputStream stream;
+
+  @Setup(Level.Trial)
+  public void buildWire() {
+    ByteWriter bytes = new ByteWriter(128);
+    bytes.writeBytes("PRI * HTTP/2.0\r\n\r\nSM\r\n\r\n".getBytes(StandardCharsets.US_ASCII));
+    FrameWriteBuffer frames = new FrameWriteBuffer(bytes);
+    frames.beginFrame(FrameType.SETTINGS, 0, 0);
+    frames.endFrame();
+    frames.beginFrame(FrameType.SETTINGS, FrameFlags.ACK, 0);
+    frames.endFrame();
+    frames.beginFrame(FrameType.PING, 0, 0);
+    bytes.writeAscii("12345678");
+    frames.endFrame();
+    frames.beginFrame(FrameType.WINDOW_UPDATE, 0, 0);
+    bytes.writeUInt31(1);
+    frames.endFrame();
+    frames.beginFrame(FrameType.GOAWAY, 0, 0);
+    bytes.writeUInt31(0);
+    bytes.writeUInt32(Http2ErrorCode.NO_ERROR.code());
+    frames.endFrame();
+    wire = new byte[bytes.length()];
+    System.arraycopy(bytes.array(), 0, wire, 0, wire.length);
+    setupConnection();
+  }
+
+  private void setupConnection() {
+    connection = new Http2Connection();
+    stream = new ResettableInputStream(wire);
+    input = new BufferedByteSource(stream, null);
+    reader = new Http2FrameReader(input);
+    sink = new CountingSink();
+    writer = new Http2FrameWriter(sink, 5_000);
+  }
+
+  @Setup(Level.Invocation)
+  public void resetConnection() {
+    stream.reset();
+    connection.reset();
+    sink.bytes = 0;
+  }
+
+  @TearDown(Level.Trial)
+  public void closeWriter() {
+    writer.close();
+  }
+
+  @Benchmark
+  public int controlLifecycle() throws Exception {
+    connection.runPrepared(input, reader, writer, RUNNING);
+    return sink.bytes;
+  }
+
+  private static final class ResettableInputStream extends InputStream {
+    private final byte[] bytes;
+    private int position;
+
+    private ResettableInputStream(byte[] bytes) {
+      this.bytes = bytes;
+    }
+
+    @Override
+    public void reset() {
+      position = 0;
+    }
+
+    @Override
+    public int read() {
+      return position == bytes.length ? -1 : bytes[position++] & 0xff;
+    }
+
+    @Override
+    public int read(byte[] target, int offset, int length) {
+      if (position == bytes.length) return -1;
+      int count = Math.min(length, bytes.length - position);
+      System.arraycopy(bytes, position, target, offset, count);
+      position += count;
+      return count;
+    }
+  }
+
+  private static final class CountingSink implements Http2FrameWriter.Sink {
+    private int bytes;
+
+    @Override
+    public void write(byte[] buffer, int offset, int length) {
+      bytes += length;
+    }
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/PerformanceGateTest.java

@@ -0,0 +1,95 @@
+package dev.relism.flash.http2;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+import java.util.Collection;
+import java.util.Map;
+import org.junit.jupiter.api.Test;
+import org.junit.jupiter.api.condition.EnabledIfSystemProperty;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.profile.GCProfiler;
+import org.openjdk.jmh.results.Result;
+import org.openjdk.jmh.results.RunResult;
+import org.openjdk.jmh.runner.Runner;
+import org.openjdk.jmh.runner.options.ChainedOptionsBuilder;
+import org.openjdk.jmh.runner.options.Options;
+import org.openjdk.jmh.runner.options.OptionsBuilder;
+import org.openjdk.jmh.runner.options.TimeValue;
+
+/** Short, forked JMH gates used by CI; full publication runs retain each benchmark's annotations. */
+@EnabledIfSystemProperty(named = "flash.performance.gates", matches = "true")
+class PerformanceGateTest {
+  private static final double ALLOCATION_NOISE_FLOOR = 0.05;
+  private static final String INCLUDE =
+      "(RequestPipelineBenchmark.parseAndRoute"
+          + "|Http2StreamBenchmark.lifecycle"
+          + "|Http2ResponseWriterBenchmark.encodeResponse"
+          + "|HpackDecoderBenchmark.decodeTypicalBrowserRequest"
+          + "|HpackEncoderBenchmark.encodeTypicalResponse"
+          + "|FrameLayerBenchmark.readValidateAndDiscard)";
+
+  @Test
+  void allocationAndLatencyBaselinesHold() throws Exception {
+    Collection<RunResult> allocationResults = new Runner(allocationOptions()).run();
+    assertFalse(allocationResults.isEmpty(), "JMH did not discover the allocation gates");
+    for (RunResult run : allocationResults) {
+      String benchmark = shortName(run.getParams().getBenchmark());
+      Result<?> allocation = run.getSecondaryResults().get("gc.alloc.rate.norm");
+      assertTrue(allocation != null, "missing allocation measurement for " + benchmark);
+      assertTrue(
+          allocation.getScore() <= ALLOCATION_NOISE_FLOOR,
+          () -> benchmark + " allocated " + allocation.getScore() + " B/op");
+    }
+
+    Collection<RunResult> latencyResults = new Runner(latencyOptions()).run();
+    assertFalse(latencyResults.isEmpty(), "JMH did not discover the latency gates");
+    for (RunResult run : latencyResults) {
+      String benchmark = shortName(run.getParams().getBenchmark());
+      Double maximumNanos = MAXIMUM_P99_NANOS.get(benchmark);
+      assertTrue(maximumNanos != null, "missing latency baseline for " + benchmark);
+      Result<?> p99 = run.getSecondaryResults().get("p0.99");
+      assertTrue(p99 != null, "missing p99 measurement for " + benchmark);
+      double score = p99.getScore();
+      assertTrue(
+          score <= maximumNanos,
+          () -> benchmark + " p99 regressed to " + score + " ns/op; gate is " + maximumNanos);
+    }
+  }
+
+  private static Options allocationOptions() {
+    return commonOptions()
+        .mode(Mode.AverageTime)
+        .addProfiler(GCProfiler.class)
+        .build();
+  }
+
+  private static Options latencyOptions() {
+    return commonOptions().mode(Mode.SampleTime).build();
+  }
+
+  private static ChainedOptionsBuilder commonOptions() {
+    return new OptionsBuilder()
+        .include(INCLUDE)
+        .warmupIterations(2)
+        .warmupTime(TimeValue.milliseconds(250))
+        .measurementIterations(3)
+        .measurementTime(TimeValue.milliseconds(350))
+        .forks(1)
+        .shouldFailOnError(true);
+  }
+
+  private static String shortName(String benchmark) {
+    return benchmark.substring(benchmark.lastIndexOf('.') + 1);
+  }
+
+  // Filled from the controlled baseline run documented in BASELINES.md, with 35% CI headroom.
+  private static final Map<String, Double> MAXIMUM_P99_NANOS =
+      Map.of(
+          "parseAndRoute", 45_000.0,
+          "lifecycle", 2_900.0,
+          "encodeResponse", 1_350.0,
+          "decodeTypicalBrowserRequest", 7_100.0,
+          "encodeTypicalResponse", 850.0,
+          "readValidateAndDiscard", 2_700.0);
+}

flash/src/jmh/java/dev/relism/flash/http2/RollingWindowCounterBenchmark.java

@@ -0,0 +1,27 @@
+package dev.relism.flash.http2;
+
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+@Fork(2)
+public class RollingWindowCounterBenchmark {
+  private final RollingWindowCounter counter = new RollingWindowCounter(10_000);
+
+  @Benchmark
+  public boolean increment() {
+    return counter.incrementExceeded(Integer.MAX_VALUE);
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/frame/FrameLayerBenchmark.java

@@ -0,0 +1,111 @@
+package dev.relism.flash.http2.frame;
+
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.transport.BufferedByteSource;
+import java.io.IOException;
+import java.io.InputStream;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/**
+ * Measures allocation and latency for reading, validating and discarding a frame and for writing a
+ * frame header. Allocation is measured with {@code -prof gc}, not inferred from inspection.
+ *
+ * <p>Uses the same hand-rolled repeating {@link InputStream} technique {@code
+ * RequestPipelineBenchmark} established: one {@link BufferedByteSource}/ {@link Http2FrameReader}
+ * pair created once per trial and reused across every invocation, matching how a real connection's
+ * demux loop owns exactly one of each for its whole lifetime, rather than paying for harness-side
+ * (re)construction inside the timed path.
+ */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class FrameLayerBenchmark {
+
+  private static final class RepeatingByteStream extends InputStream {
+    private final byte[] template;
+    private int pos;
+
+    RepeatingByteStream(byte[] template) {
+      this.template = template;
+    }
+
+    @Override
+    public int read() {
+      byte b = template[pos];
+      pos = (pos + 1) % template.length;
+      return b & 0xFF;
+    }
+
+    @Override
+    public int read(byte[] dst, int off, int len) {
+      for (int i = 0; i < len; i++) {
+        dst[off + i] = template[pos];
+        pos = (pos + 1) % template.length;
+      }
+      return len;
+    }
+  }
+
+  // ── Read + validate ──────────────────────────────────────────────────────
+
+  private Http2FrameReader reader;
+
+  @Setup(Level.Trial)
+  public void setupReader() {
+    FrameWriteBuffer out = new FrameWriteBuffer(new ByteWriter(64));
+    out.beginFrame(FrameType.HEADERS, FrameFlags.END_HEADERS, 1);
+    byte[] payload = new byte[48];
+    for (int i = 0; i < payload.length; i++) payload[i] = (byte) i;
+    out.writer().writeBytes(payload);
+    out.endFrame();
+    byte[] template = new byte[out.writer().length()];
+    System.arraycopy(out.writer().array(), 0, template, 0, template.length);
+
+    BufferedByteSource src = new BufferedByteSource(new RepeatingByteStream(template), null);
+    reader = new Http2FrameReader(src);
+  }
+
+  @Benchmark
+  public int readValidateAndDiscard() throws IOException {
+    FrameHeader header = reader.readFrame();
+    FrameValidator.validate(header, false);
+    int checksum = header.buffer()[header.payloadOffset()];
+    reader.consumeFrame();
+    return checksum;
+  }
+
+  // ── Write ────────────────────────────────────────────────────────────────
+
+  private FrameWriteBuffer writeBuffer;
+  private byte[] writePayload;
+
+  @Setup(Level.Trial)
+  public void setupWriter() {
+    writeBuffer = new FrameWriteBuffer(new ByteWriter(64));
+    writePayload = new byte[48];
+    for (int i = 0; i < writePayload.length; i++) writePayload[i] = (byte) i;
+  }
+
+  @Benchmark
+  public int writeFrame() {
+    writeBuffer.writer().reset();
+    writeBuffer.beginFrame(FrameType.HEADERS, FrameFlags.END_HEADERS, 1);
+    writeBuffer.writer().writeBytes(writePayload);
+    writeBuffer.endFrame();
+    return writeBuffer.writer().length();
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/frame/FrameWriterBenchmark.java

@@ -0,0 +1,353 @@
+package dev.relism.flash.http2.frame;
+
+import java.util.Arrays;
+import java.util.Set;
+import java.util.concurrent.ConcurrentHashMap;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+import java.util.concurrent.Future;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicLong;
+import java.util.concurrent.locks.LockSupport;
+import java.util.concurrent.locks.ReentrantLock;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Param;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.TearDown;
+import org.openjdk.jmh.annotations.Warmup;
+
+/**
+ * Compares three writer designs at N ∈ {1, 2, 4, 8, 16, 64} concurrent <b>virtual-thread</b>
+ * writers:
+ *
+ * <ul>
+ *   <li>{@code trylock_mpsc} — the design that ships as {@link Http2FrameWriter}: {@code tryLock()}
+ *       fast path, intrusive MPSC fallback.
+ *   <li>{@code plain_lock} — every write blocks on {@code ReentrantLock#lock()}, unconditionally.
+ *   <li>{@code dedicated_thread} — every write hands off to a single dedicated platform thread via
+ *       the same {@link IntrusiveMpscQueue}, parked/unparked (never a busy poll).
+ * </ul>
+ *
+ * <h3>Why this benchmark drives its own concurrency instead of JMH's {@code @Threads}</h3>
+ *
+ * {@code @Threads} requires a compile-time constant, not a {@code @Param}-swept value, and JMH's
+ * thread pool is platform threads, not virtual threads — the exact scheduling behaviour under test.
+ * Each {@code @Benchmark} invocation therefore spawns {@link #threads} virtual threads itself, has
+ * them race a fixed burst of writes to a counting no-op sink, and reports the burst's wall-clock
+ * rate; JMH still owns fork/warmup/measurement-iteration control and (via {@code -prof gc}) the
+ * zero-allocation verification.
+ *
+ * <h3>Why {@code runBurst} waits on a write counter, not just thread completion</h3>
+ *
+ * {@code write()} does not mean "already on the wire" for every design: the shipped design's
+ * contended path, and the dedicated-thread design's handoff, can both return once the frame is
+ * merely *queued*. Timing only "how long until every producer's {@code write()} call returned"
+ * would therefore measure submission speed, not completion speed, and would flatter exactly the
+ * designs that most aggressively defer work — the opposite of a fair comparison. Every harness here
+ * writes through {@link CountingSink} and {@link #runBurst} waits for its counter to reach the
+ * expected total before returning, so the timed interval always covers real completion.
+ *
+ * <p>Per-write latency percentiles are computed by hand from {@code System.nanoTime()} samples
+ * collected during the burst (JMH's own {@code Mode.SampleTime} does not fit a custom-concurrency
+ * benchmark method) and printed once per (design, threads) combination — see {@code WRITER.md} for
+ * the recorded results and the gate decision.
+ *
+ * <p>Run: {@code mvn -Pjmh -pl flash test-compile} then {@code java -cp
+ * flash/target/test-classes:flash/target/classes:$(mvn -Pjmh -pl flash dependency:build-classpath
+ * -Dmdep.outputFile=/dev/stdout -q) org.openjdk.jmh.Main FrameWriterBenchmark -prof gc}.
+ */
+@State(Scope.Benchmark)
+@BenchmarkMode(Mode.Throughput)
+@OutputTimeUnit(TimeUnit.SECONDS)
+@Fork(1)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class FrameWriterBenchmark {
+
+  private static final int FRAMES_PER_THREAD = 4000;
+  private static final int FRAME_SIZE = 512;
+
+  @Param({"trylock_mpsc", "plain_lock", "dedicated_thread", "raw_unsynchronized"})
+  public String design;
+
+  @Param({"1", "2", "4", "8", "16", "64"})
+  public int threads;
+
+  private DesignHarness harness;
+  private byte[] payload;
+
+  @Setup(Level.Trial)
+  public void setup() {
+    payload = new byte[FRAME_SIZE];
+    harness =
+        switch (design) {
+          case "trylock_mpsc" -> new TryLockMpscHarness();
+          case "plain_lock" -> new PlainLockHarness();
+          case "dedicated_thread" -> new DedicatedThreadHarness();
+          case "raw_unsynchronized" -> new RawUnsynchronizedHarness();
+          default -> throw new IllegalStateException("unknown design: " + design);
+        };
+  }
+
+  @TearDown(Level.Trial)
+  public void teardown() {
+    harness.shutdown();
+  }
+
+  /**
+   * One "operation" here is a full burst: {@link #threads} virtual threads each writing {@link
+   * #FRAMES_PER_THREAD} frames. Reported ops/sec must be multiplied by {@code threads *
+   * FRAMES_PER_THREAD} to get frames/sec — done during result analysis, not via
+   * {@code @OperationsPerInvocation} (which requires a compile-time constant and cannot vary with
+   * the {@code threads} @Param).
+   */
+  @Benchmark
+  public void burst() throws Exception {
+    harness.runBurst(threads, FRAMES_PER_THREAD, payload);
+  }
+
+  // ── Harness abstraction and the three designs under comparison ─────────────
+
+  private interface DesignHarness {
+    void runBurst(int threads, int framesPerThread, byte[] payload) throws Exception;
+
+    void shutdown();
+  }
+
+  /**
+   * Discards everything (isolating the writer designs from real socket variance) but counts every
+   * completed write, so callers can wait for true completion rather than mere submission — see the
+   * class Javadoc.
+   */
+  private static final class CountingSink implements Http2FrameWriter.Sink {
+    final AtomicLong count = new AtomicLong();
+
+    @Override
+    public void write(byte[] buf, int off, int len) {
+      count.incrementAndGet();
+    }
+  }
+
+  private static final class BenchIntent implements WriteIntent {
+    final byte[] buf;
+    WriteIntent next;
+
+    BenchIntent(byte[] buf) {
+      this.buf = buf;
+    }
+
+    @Override
+    public byte[] buffer() {
+      return buf;
+    }
+
+    @Override
+    public int offset() {
+      return 0;
+    }
+
+    @Override
+    public int length() {
+      return buf.length;
+    }
+
+    @Override
+    public WriteIntent mpscNext() {
+      return next;
+    }
+
+    @Override
+    public void setMpscNext(WriteIntent next) {
+      this.next = next;
+    }
+  }
+
+  private interface ThrowingConsumer<T> {
+    void accept(T t) throws Exception;
+  }
+
+  /**
+   * Spawns {@code threadCount} virtual threads, has each write {@code framesPerThread} fresh {@link
+   * BenchIntent}s (one per write — matches production usage, where a stream's scratch buffer holds
+   * exactly one in-flight frame at a time), records per-write latency samples, then blocks until
+   * {@code sink}'s counter reflects every one of them actually written.
+   */
+  private static void race(
+      int threadCount, int framesPerThread, CountingSink sink, ThrowingConsumer<WriteIntent> write)
+      throws Exception {
+    long target = sink.count.get() + (long) threadCount * framesPerThread;
+    byte[] payload = new byte[FRAME_SIZE];
+    long[][] samplesByThread = new long[threadCount][framesPerThread];
+    try (ExecutorService exec = Executors.newVirtualThreadPerTaskExecutor()) {
+      Future<?>[] futures = new Future<?>[threadCount];
+      for (int t = 0; t < threadCount; t++) {
+        int idx = t;
+        futures[t] =
+            exec.submit(
+                () -> {
+                  long[] samples = samplesByThread[idx];
+                  for (int i = 0; i < framesPerThread; i++) {
+                    BenchIntent intent = new BenchIntent(payload);
+                    long start = System.nanoTime();
+                    try {
+                      write.accept(intent);
+                    } catch (Exception e) {
+                      throw new RuntimeException(e);
+                    }
+                    samples[i] = System.nanoTime() - start;
+                  }
+                });
+      }
+      for (Future<?> f : futures) f.get();
+    }
+    while (sink.count.get() < target) {
+      Thread.onSpinWait();
+    }
+    LatencyReport.recordAndMaybePrint(samplesByThread);
+  }
+
+  /**
+   * Prints p50/p99/p999 to stdout once per thread-count actually exercised, from the first burst
+   * observed for it — cheap, and avoids flooding the JMH log with one line per measurement
+   * iteration.
+   */
+  private static final class LatencyReport {
+    private static final Set<String> PRINTED = ConcurrentHashMap.newKeySet();
+
+    static void recordAndMaybePrint(long[][] samplesByThread) {
+      String key = samplesByThread.length + "t";
+      if (!PRINTED.add(key)) return;
+
+      int total = 0;
+      for (long[] s : samplesByThread) total += s.length;
+      long[] all = new long[total];
+      int pos = 0;
+      for (long[] s : samplesByThread) {
+        System.arraycopy(s, 0, all, pos, s.length);
+        pos += s.length;
+      }
+      Arrays.sort(all);
+      long p50 = all[(int) (all.length * 0.50)];
+      long p99 = all[(int) (all.length * 0.99)];
+      long p999 = all[Math.min(all.length - 1, (int) (all.length * 0.999))];
+      System.out.printf(
+          "[latency threads=%d] p50=%.1fus p99=%.1fus p999=%.1fus (n=%d)%n",
+          samplesByThread.length, p50 / 1000.0, p99 / 1000.0, p999 / 1000.0, all.length);
+    }
+  }
+
+  // ── Baseline: no synchronization at all ─────────────────────────────────────
+  // Not a candidate design (concurrent writers would tear each other's frames) — exists
+  // purely to establish "what a write costs with zero coordination overhead" for the N=1
+  // gate criterion ("per-frame overhead versus a raw unsynchronized write is within 50 ns").
+  // At N=1 there genuinely is no concurrent writer, so the missing safety is moot there.
+
+  private static final class RawUnsynchronizedHarness implements DesignHarness {
+    private final CountingSink sink = new CountingSink();
+
+    @Override
+    public void runBurst(int threads, int framesPerThread, byte[] payload) throws Exception {
+      race(
+          threads,
+          framesPerThread,
+          sink,
+          intent -> sink.write(intent.buffer(), intent.offset(), intent.length()));
+    }
+
+    @Override
+    public void shutdown() {}
+  }
+
+  // ── Design (a): plain lock ──────────────────────────────────────────────────
+
+  private static final class PlainLockHarness implements DesignHarness {
+    private final CountingSink sink = new CountingSink();
+    private final ReentrantLock lock = new ReentrantLock();
+
+    @Override
+    public void runBurst(int threads, int framesPerThread, byte[] payload) throws Exception {
+      race(
+          threads,
+          framesPerThread,
+          sink,
+          intent -> {
+            lock.lock();
+            try {
+              sink.write(intent.buffer(), intent.offset(), intent.length());
+            } finally {
+              lock.unlock();
+            }
+          });
+    }
+
+    @Override
+    public void shutdown() {}
+  }
+
+  // ── Design (b): tryLock + intrusive MPSC — the shipped design ──────────────
+
+  private static final class TryLockMpscHarness implements DesignHarness {
+    private final CountingSink sink = new CountingSink();
+    private final Http2FrameWriter writer = new Http2FrameWriter(sink, 30_000);
+
+    @Override
+    public void runBurst(int threads, int framesPerThread, byte[] payload) throws Exception {
+      race(threads, framesPerThread, sink, writer::write);
+    }
+
+    @Override
+    public void shutdown() {
+      writer.close();
+    }
+  }
+
+  // ── Design (c): always hand off to one dedicated writer thread ─────────────
+
+  private static final class DedicatedThreadHarness implements DesignHarness {
+    private final CountingSink sink = new CountingSink();
+    private final IntrusiveMpscQueue queue = new IntrusiveMpscQueue();
+    private final Thread writerThread;
+    private volatile boolean running = true;
+
+    DedicatedThreadHarness() {
+      this.writerThread = Thread.ofPlatform().name("bench-dedicated-writer").start(this::loop);
+    }
+
+    private void loop() {
+      while (running) {
+        WriteIntent intent = queue.poll();
+        if (intent == null) {
+          LockSupport.park();
+          continue;
+        }
+        sink.write(intent.buffer(), intent.offset(), intent.length());
+      }
+    }
+
+    @Override
+    public void runBurst(int threads, int framesPerThread, byte[] payload) throws Exception {
+      race(
+          threads,
+          framesPerThread,
+          sink,
+          intent -> {
+            queue.offer(intent);
+            LockSupport.unpark(writerThread);
+          });
+    }
+
+    @Override
+    public void shutdown() {
+      running = false;
+      writerThread.interrupt();
+    }
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/hpack/HpackDecoderBenchmark.java

@@ -0,0 +1,62 @@
+package dev.relism.flash.http2.hpack;
+
+import java.util.concurrent.TimeUnit;
+import dev.relism.flash.bytes.ByteWriter;
+import java.nio.charset.StandardCharsets;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/** Measures steady-state decoding into reusable per-stream storage. */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class HpackDecoderBenchmark {
+  private final HpackDecoder decoder = new HpackDecoder();
+  private final HpackHeaderBlock headers = new HpackHeaderBlock();
+  private final byte[] block = {(byte) 0x82, (byte) 0x87, (byte) 0x84, (byte) 0x88};
+  private byte[] browserBlock;
+
+  @Setup(Level.Trial)
+  public void setupTypicalBlock() {
+    ByteWriter encoded = new ByteWriter(256);
+    HpackEncoder.writeIndexed(encoded, 2);
+    HpackEncoder.writeIndexed(encoded, 7);
+    HpackEncoder.writeLiteralWithNameIndex(
+        encoded, 4, "/products?category=books".getBytes(StandardCharsets.US_ASCII), true);
+    HpackEncoder.writeLiteralWithNameIndex(
+        encoded, 1, "shop.example.com".getBytes(StandardCharsets.US_ASCII), true);
+    HpackEncoder.writeLiteralWithNameIndex(
+        encoded, 19, "text/html,application/xhtml+xml".getBytes(StandardCharsets.US_ASCII), true);
+    HpackEncoder.writeLiteralWithNameIndex(
+        encoded, 16, "gzip, deflate".getBytes(StandardCharsets.US_ASCII), true);
+    HpackEncoder.writeLiteralWithNameIndex(
+        encoded, 55, "Mozilla/5.0 benchmark".getBytes(StandardCharsets.US_ASCII), true);
+    browserBlock = java.util.Arrays.copyOf(encoded.array(), encoded.length());
+  }
+
+  @Benchmark
+  public int decodeStaticRequest() {
+    headers.reset();
+    decoder.decode(block, 0, block.length, headers);
+    return headers.count();
+  }
+
+  @Benchmark
+  public int decodeTypicalBrowserRequest() {
+    headers.reset();
+    decoder.decode(browserBlock, 0, browserBlock.length, headers);
+    return headers.count();
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/hpack/HpackEncoderBenchmark.java

@@ -0,0 +1,43 @@
+package dev.relism.flash.http2.hpack;
+
+import dev.relism.flash.bytes.ByteWriter;
+import java.nio.charset.StandardCharsets;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/** Measures a representative stateless response header block. */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class HpackEncoderBenchmark {
+  private static final byte[] CONTENT_LENGTH = "1024".getBytes(StandardCharsets.US_ASCII);
+  private static final byte[] CONTENT_TYPE = "application/json".getBytes(StandardCharsets.US_ASCII);
+  private static final byte[] CACHE_CONTROL = "no-cache".getBytes(StandardCharsets.US_ASCII);
+  private static final byte[] ETAG_NAME = "etag".getBytes(StandardCharsets.US_ASCII);
+  private static final byte[] ETAG = "\"abc123\"".getBytes(StandardCharsets.US_ASCII);
+  private static final byte[] SERVER = "Flash".getBytes(StandardCharsets.US_ASCII);
+  private final ByteWriter output = new ByteWriter(128);
+
+  @Benchmark
+  public int encodeTypicalResponse() {
+    output.reset();
+    HpackEncoder.writeIndexed(output, 8);
+    HpackEncoder.writeLiteralWithNameIndex(output, 31, CONTENT_TYPE, true);
+    HpackEncoder.writeLiteralWithNameIndex(output, 28, CONTENT_LENGTH, false);
+    HpackEncoder.writeLiteralWithNameIndex(output, 24, CACHE_CONTROL, true);
+    HpackEncoder.writeLiteralWithNameIndex(output, 51, SERVER, true);
+    HpackEncoder.writeLiteral(output, ETAG_NAME, ETAG);
+    return output.length();
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/message/Http2BodyBenchmark.java

@@ -0,0 +1,122 @@
+package dev.relism.flash.http2.message;
+
+import dev.relism.flash.http.ContentType;
+import dev.relism.flash.models.RequestBody;
+import dev.relism.flash.models.Response;
+import java.io.IOException;
+import java.io.InputStream;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Warmup(iterations = 3)
+@Measurement(iterations = 5)
+@Fork(2)
+@State(Scope.Thread)
+public class Http2BodyBenchmark {
+  private static final Http2RequestBody.ConsumptionListener NOOP = bytes -> {};
+
+  private final byte[] payload = new byte[1024];
+  private final byte[] streamingPayload = new byte[1024 * 1024];
+  private final byte[] target = new byte[1024];
+  private DataBufferPool pool;
+  private Http2RequestBody source;
+  private RequestBody body;
+  private Response response;
+  private Http2ResponseWriter responseWriter;
+  private ResettableInputStream responseSource;
+  private ResettableInputStream largeResponseSource;
+
+  @Setup(Level.Trial)
+  public void setup() throws IOException {
+    pool = new DataBufferPool(16_384, 1);
+    source = new Http2RequestBody(pool);
+    body = new RequestBody();
+    response = new Response(200, ContentType.BINARY);
+    responseWriter = new Http2ResponseWriter();
+    responseSource = new ResettableInputStream(payload);
+    largeResponseSource = new ResettableInputStream(streamingPayload);
+    source.begin(-1, false, NOOP);
+    source.offer(1, payload, 0, payload.length, payload.length);
+    source.finish(1);
+    source.read(target);
+  }
+
+  @Benchmark
+  public byte[] inlineBytes() {
+    source.begin(payload.length, true, NOOP);
+    source.offer(1, payload, 0, payload.length, payload.length);
+    source.finish(1);
+    body.reset(source, payload.length, null, 0, 0);
+    return body.bytes();
+  }
+
+  @Benchmark
+  public int streamingRead() throws IOException {
+    source.begin(-1, false, NOOP);
+    source.offer(1, payload, 0, payload.length, payload.length);
+    source.finish(1);
+    return source.read(target, 0, target.length);
+  }
+
+  @Benchmark
+  public int streamingResponseFrame() throws IOException {
+    responseSource.rewind();
+    response.reset(200, ContentType.BINARY).stream(responseSource, payload.length);
+    responseWriter.startFlowControlled(
+        response, 1, false, false, true, false, false, 16_384, 32_768, 16_384);
+    return responseWriter.length();
+  }
+
+  @Benchmark
+  public int streamingResponseOneMiB() throws IOException {
+    largeResponseSource.rewind();
+    response.reset(200, ContentType.BINARY).stream(largeResponseSource, streamingPayload.length);
+    responseWriter.startFlowControlled(
+        response, 1, false, false, true, false, false, 16_384, 32_768, 16_384);
+    int wireBytes = responseWriter.length();
+    while (!responseWriter.finished()) {
+      responseWriter.resume(16_384, 16_384);
+      wireBytes += responseWriter.length();
+    }
+    return wireBytes;
+  }
+
+  private static final class ResettableInputStream extends InputStream {
+    private final byte[] source;
+    private int position;
+
+    ResettableInputStream(byte[] source) {
+      this.source = source;
+    }
+
+    void rewind() {
+      position = 0;
+    }
+
+    @Override
+    public int read() {
+      return position == source.length ? -1 : source[position++] & 0xff;
+    }
+
+    @Override
+    public int read(byte[] target, int offset, int length) {
+      if (position == source.length) return -1;
+      int count = Math.min(length, source.length - position);
+      System.arraycopy(source, position, target, offset, count);
+      position += count;
+      return count;
+    }
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/message/Http2ResponseWriterBenchmark.java

@@ -0,0 +1,44 @@
+package dev.relism.flash.http2.message;
+
+import dev.relism.flash.http.ContentType;
+import dev.relism.flash.models.PreEncodedHeader;
+import dev.relism.flash.models.Response;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/** Measures the steady-state allocation cost of a representative fixed HTTP/2 response. */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class Http2ResponseWriterBenchmark {
+  private Http2ResponseWriter writer;
+  private Response response;
+
+  @Setup
+  public void setup() {
+    writer = new Http2ResponseWriter();
+    response =
+        new Response(200, "hello", ContentType.JSON)
+            .header(new PreEncodedHeader("cache-control", "no-store"))
+            .header(new PreEncodedHeader("x-trace", "abc123"));
+    writer.prepare(response, 1, false, true, true, false, false, 16_384, 16_384, 65_535);
+  }
+
+  @Benchmark
+  public int encodeResponse() {
+    writer.prepare(response, 1, false, true, true, false, false, 16_384, 16_384, 65_535);
+    return writer.length();
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/message/Http2TuningBenchmark.java

@@ -0,0 +1,117 @@
+package dev.relism.flash.http2.message;
+
+import dev.relism.flash.http.ContentType;
+import dev.relism.flash.models.PreEncodedHeader;
+import dev.relism.flash.models.Response;
+import java.io.IOException;
+import java.io.InputStream;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Param;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/** Factorial measurements for the response knobs considered during tuning. */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class Http2TuningBenchmark {
+  @Param({"false", "true"})
+  public boolean huffmanDynamicValues;
+
+  @Param({"16384", "65536", "1048576"})
+  public int maxFrameSize;
+
+  private final byte[] largeBody = new byte[1024 * 1024];
+  private Http2ResponseWriter writer;
+  private Response response;
+  private Response streamingResponse;
+  private ResettableInputStream source;
+
+  @Setup(Level.Trial)
+  public void setup() {
+    writer = new Http2ResponseWriter();
+    response =
+        new Response(200, "hello", ContentType.JSON)
+            .header(new PreEncodedHeader("cache-control", "private, max-age=60"))
+            .header(new PreEncodedHeader("x-request-id", "d7bca219-6dd4-4ef0-a881-f21931e249c7"));
+    source = new ResettableInputStream(largeBody);
+    streamingResponse = new Response(200, ContentType.BINARY).stream(source, largeBody.length);
+  }
+
+  @Benchmark
+  public int encodeResponseHeaders() {
+    writer.prepare(
+        response,
+        1,
+        false,
+        true,
+        true,
+        huffmanDynamicValues,
+        false,
+        maxFrameSize,
+        32_768,
+        65_535);
+    return writer.length();
+  }
+
+  @Benchmark
+  public int streamOneMiB() throws IOException {
+    source.rewind();
+    writer.startFlowControlled(
+        streamingResponse,
+        1,
+        false,
+        false,
+        true,
+        huffmanDynamicValues,
+        false,
+        maxFrameSize,
+        32_768,
+        maxFrameSize);
+    int wireBytes = writer.length();
+    while (!writer.finished()) {
+      writer.resume(maxFrameSize, maxFrameSize);
+      wireBytes += writer.length();
+    }
+    return wireBytes;
+  }
+
+  private static final class ResettableInputStream extends InputStream {
+    private final byte[] bytes;
+    private int position;
+
+    private ResettableInputStream(byte[] bytes) {
+      this.bytes = bytes;
+    }
+
+    private void rewind() {
+      position = 0;
+    }
+
+    @Override
+    public int read() {
+      return position == bytes.length ? -1 : bytes[position++] & 0xff;
+    }
+
+    @Override
+    public int read(byte[] target, int offset, int length) {
+      if (position == bytes.length) return -1;
+      int count = Math.min(length, bytes.length - position);
+      System.arraycopy(bytes, position, target, offset, count);
+      position += count;
+      return count;
+    }
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/stream/Http2MultiplexingBenchmark.java

@@ -0,0 +1,77 @@
+package dev.relism.flash.http2.stream;
+
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http2.hpack.HpackDecoder;
+import dev.relism.flash.http2.hpack.HpackEncoder;
+import java.nio.charset.StandardCharsets;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Param;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/** Measures one response pass across a connection with N simultaneously live request streams. */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class Http2MultiplexingBenchmark {
+  private static final byte[] BODY = "ok".getBytes(StandardCharsets.US_ASCII);
+
+  @Param({"1", "8", "64", "256"})
+  public int liveStreams;
+
+  private Http2Stream[] streams;
+
+  @Setup(Level.Trial)
+  public void setup() {
+    Http2StreamTable table = new Http2StreamTable(liveStreams);
+    streams = new Http2Stream[liveStreams];
+    ByteWriter encoded = new ByteWriter(64);
+    HpackEncoder.writeIndexed(encoded, 2);
+    HpackEncoder.writeIndexed(encoded, 7);
+    HpackEncoder.writeLiteralWithNameIndex(
+        encoded, 4, "/get".getBytes(StandardCharsets.US_ASCII), false);
+    HpackEncoder.writeLiteralWithNameIndex(
+        encoded, 1, "localhost".getBytes(StandardCharsets.US_ASCII), false);
+    HpackDecoder decoder = new HpackDecoder();
+    for (int i = 0; i < streams.length; i++) {
+      Http2Stream stream = table.acquire(i * 2 + 1);
+      decoder.decode(encoded.array(), 0, encoded.length(), stream.headerBlock());
+      stream.assembleRequest(null, null);
+      streams[i] = stream;
+    }
+  }
+
+  @Benchmark
+  public int encodeAllLiveStreamResponses() {
+    int wireBytes = 0;
+    for (Http2Stream stream : streams) {
+      stream
+          .responseWriter()
+          .prepare(
+              stream.resetResponse().body(BODY),
+              stream.id(),
+              false,
+              false,
+              true,
+              false,
+              false,
+              16_384,
+              32_768,
+              65_535);
+      wireBytes += stream.responseWriter().length();
+    }
+    return wireBytes;
+  }
+}

flash/src/jmh/java/dev/relism/flash/http2/stream/Http2StreamBenchmark.java

@@ -0,0 +1,108 @@
+package dev.relism.flash.http2.stream;
+
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http2.hpack.HpackDecoder;
+import dev.relism.flash.http2.hpack.HpackEncoder;
+import dev.relism.flash.models.Response;
+import java.nio.charset.StandardCharsets;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/** Measures the pooled HPACK-decode, request-assembly and fixed-response stream lifecycle. */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class Http2StreamBenchmark {
+  private static final byte[] BODY = "pong".getBytes(StandardCharsets.US_ASCII);
+  private static final byte[] POST_BODY = new byte[1024];
+
+  private Http2StreamTable streams;
+  private HpackDecoder decoder;
+  private byte[] requestBlock;
+  private int requestLength;
+  private byte[] postBlock;
+  private int postLength;
+
+  @Setup
+  public void setup() {
+    streams = new Http2StreamTable(1);
+    decoder = new HpackDecoder();
+    ByteWriter block = new ByteWriter(64);
+    HpackEncoder.writeIndexed(block, 2);
+    HpackEncoder.writeIndexed(block, 7);
+    HpackEncoder.writeLiteralWithNameIndex(
+        block, 4, "/ping".getBytes(StandardCharsets.US_ASCII), false);
+    HpackEncoder.writeLiteralWithNameIndex(
+        block, 1, "localhost".getBytes(StandardCharsets.US_ASCII), false);
+    requestBlock = block.array();
+    requestLength = block.length();
+    ByteWriter post = new ByteWriter(96);
+    HpackEncoder.writeIndexed(post, 3);
+    HpackEncoder.writeIndexed(post, 7);
+    HpackEncoder.writeLiteralWithNameIndex(
+        post, 4, "/ping".getBytes(StandardCharsets.US_ASCII), false);
+    HpackEncoder.writeLiteralWithNameIndex(
+        post, 1, "localhost".getBytes(StandardCharsets.US_ASCII), false);
+    HpackEncoder.writeLiteralWithNameIndex(
+        post, 28, "1024".getBytes(StandardCharsets.US_ASCII), false);
+    postBlock = post.array();
+    postLength = post.length();
+    lifecycle();
+    postOneKiB();
+  }
+
+  @Benchmark
+  public int lifecycle() {
+    Http2Stream stream = streams.acquire(1);
+    decoder.decode(requestBlock, 0, requestLength, stream.headerBlock());
+    stream.assembleRequest(null, null);
+    Response response = stream.resetResponse().body(BODY);
+    stream
+        .responseWriter()
+        .prepare(response, 1, false, false, true, false, false, 16_384, 32_768, 65_535);
+    int bytes = stream.responseWriter().length();
+    streams.remove(1);
+    streams.release(stream);
+    return bytes;
+  }
+
+  /** Unary request shape: HPACK decode, one 1 KiB DATA payload, assembly and fixed response. */
+  @Benchmark
+  public int postOneKiB() {
+    Http2Stream stream = streams.acquire(1);
+    decoder.decode(postBlock, 0, postLength, stream.headerBlock());
+    stream.prepareRequestBody(null, false);
+    stream.receiveData(POST_BODY, 0, POST_BODY.length, POST_BODY.length);
+    stream.finishRequestBody();
+    stream.assembleRequest(null, null);
+    stream
+        .responseWriter()
+        .prepare(
+            stream.resetResponse().body(BODY),
+            1,
+            false,
+            false,
+            true,
+            false,
+            false,
+            16_384,
+            32_768,
+            65_535);
+    int bytes = stream.responseWriter().length();
+    streams.remove(1);
+    streams.release(stream);
+    return bytes;
+  }
+}

flash/src/jmh/java/dev/relism/flash/routing/routers/fastpathrouter/FastPathRouterBenchmark.java

@@ -0,0 +1,119 @@
+package dev.relism.flash.routing.routers.fastpathrouter;
+
+import dev.relism.flash.http.HttpMethod;
+import dev.relism.flash.models.Request;
+import dev.relism.flash.models.RequestHandler;
+import dev.relism.flash.models.SimpleHandler;
+import dev.relism.fpr.core.internal.runtime.ByteCompare;
+import java.nio.charset.StandardCharsets;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Warmup;
+
+/**
+ * Two related but distinct measurements of router and byte-comparison performance.
+ *
+ * <p><b>{@code router_*}</b> exercises the shipped {@link FastPathRouterImpl#route} end to end,
+ * including lazy-compiled route-table lookup, {@link FastPathRouterImpl.RouteScratch} reuse and
+ * path-parameter extraction.
+ *
+ * <p><b>{@code byteCompare_*}</b> directly compares {@code ByteCompare.equals} with its
+ * word-at-a-time path enabled and disabled over representative array-backed content. This is
+ * separate because router matching uses the composite, non-array-backed {@link
+ * FastPathViews.MethodPathByteView}, so the router measurements cannot expose the word path's
+ * effect.
+ */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Fork(2)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 5, time = 1)
+public class FastPathRouterBenchmark {
+
+  // ── router_*: the real, shipped router, end to end ──────────────────────
+
+  private FastPathRouterImpl router;
+  private Object scratch;
+  private Request staticRequest;
+  private Request paramRequest;
+
+  @Setup(Level.Trial)
+  public void setupRouter() {
+    router = new FastPathRouterImpl();
+    RequestHandler h = new SimpleHandler((req, res) -> "ok");
+    router.doRegister(HttpMethod.GET, "/health", h, new dev.relism.flash.routing.Middleware[0]);
+    router.doRegister(HttpMethod.GET, "/users/{id}", h, new dev.relism.flash.routing.Middleware[0]);
+    router.doRegister(
+        HttpMethod.GET,
+        "/users/{id}/posts/{postId}",
+        h,
+        new dev.relism.flash.routing.Middleware[0]);
+    router.doRegister(HttpMethod.POST, "/users", h, new dev.relism.flash.routing.Middleware[0]);
+    router.doRegister(
+        HttpMethod.GET,
+        "/api/v1/products/{category}/{id}",
+        h,
+        new dev.relism.flash.routing.Middleware[0]);
+    router.compile();
+    scratch = router.newScratch();
+
+    staticRequest = mockRequest(HttpMethod.GET, "/health");
+    paramRequest = mockRequest(HttpMethod.GET, "/users/12345/posts/67890");
+  }
+
+  @Benchmark
+  public RequestHandler router_staticRoute() {
+    return router.route(staticRequest, scratch);
+  }
+
+  @Benchmark
+  public RequestHandler router_parametricRoute() {
+    return router.route(paramRequest, scratch);
+  }
+
+  private static Request mockRequest(HttpMethod method, String path) {
+    byte[] bytes = path.getBytes(StandardCharsets.UTF_8);
+    FastPathViews.RequestByteView pathView =
+        new FastPathViews.RequestByteView(bytes, 0, bytes.length);
+    dev.relism.flash.models.RequestLine line =
+        new dev.relism.flash.models.RequestLine(
+            method,
+            pathView,
+            null,
+            new FastPathViews.RequestByteView("HTTP/1.1".getBytes(StandardCharsets.UTF_8), 0, 8),
+            new dev.relism.flash.models.Http1HeaderMap());
+    return new Request(line, new byte[0]);
+  }
+
+  // ── byteCompare_*: word-at-a-time comparison in isolation ──────────────
+
+  private FastPathViews.RequestByteView cmpView;
+  private byte[] cmpOther;
+
+  @Setup(Level.Trial)
+  public void setupByteCompare() {
+    byte[] content = "/api/v1/products/electronics/00012345".getBytes(StandardCharsets.US_ASCII);
+    cmpView = new FastPathViews.RequestByteView(content, 0, content.length);
+    cmpOther = content.clone();
+  }
+
+  @Benchmark
+  public boolean byteCompare_longPath() {
+    return ByteCompare.equals(cmpView, 0, cmpOther, 0, cmpOther.length, true);
+  }
+
+  @Benchmark
+  public boolean byteCompare_byteAtATime() {
+    return ByteCompare.equals(cmpView, 0, cmpOther, 0, cmpOther.length, false);
+  }
+}

flash/src/main/java/dev/relism/flash/ChunkedInputStream.java

@@ -1,24 +1,49 @@
 package dev.relism.flash;
 
-import java.io.ByteArrayInputStream;
+import dev.relism.flash.exceptions.MalformedRequestException;
+import dev.relism.flash.http.Http1Limits;
+import dev.relism.flash.models.BodyCompletion;
+import dev.relism.flash.models.MutableHeaderMap;
+import dev.relism.flash.transport.BufferedByteSource;
+
 import java.io.IOException;
 import java.io.InputStream;
-import java.io.SequenceInputStream;
 
 /**
  * De-chunking {@link InputStream} for HTTP/1.1 {@code Transfer-Encoding: chunked} request bodies.
  * Handles pre-buffered bytes from the header read-ahead, chunk framing, and trailer consumption.
  * Returns -1 at end of the final chunk; the underlying socket is left positioned for the next request.
+ *
+ * instead of the raw, unbuffered socket stream. Chunk-size digits, the trailing CRLF after each
+ * chunk, and trailer lines are all read one byte at a time by design (the framing is
+ * byte-oriented) — that used to mean one {@code read(2)} syscall per byte on the raw socket;
+ * against {@link BufferedByteSource} it is a read from an already-filled in-memory buffer.
+ * The header-parser's read-ahead bytes are handed to {@code src} via
+ * {@link BufferedByteSource#prependOnce}, replacing the {@code SequenceInputStream}/
+ * {@code ByteArrayInputStream} pair the previous implementation allocated per chunked request.
  */
-final class ChunkedInputStream extends InputStream {
-    private final InputStream src;
+final class ChunkedInputStream extends InputStream implements BodyCompletion {
+    private final BufferedByteSource src;
     private int     chunkRemaining = 0;
     private boolean done           = false;
+    private int     chunksSeen     = 0;
+    private final MutableHeaderMap trailers;
+    private final byte[] trailerLine = new byte[Http1Limits.MAX_HEADER_VALUE_LENGTH];
 
-    ChunkedInputStream(InputStream socket, byte[] preBuf, int preBufOff, int preBufLen) {
-        src = preBufLen > 0
-                ? new SequenceInputStream(new ByteArrayInputStream(preBuf, preBufOff, preBufLen), socket)
-                : socket;
+    ChunkedInputStream(BufferedByteSource src, byte[] preBuf, int preBufOff, int preBufLen,
+                       MutableHeaderMap trailers) {
+        this.src = src;
+        this.trailers = trailers;
+        if (preBufLen > 0) src.prependOnce(preBuf, preBufOff, preBufLen);
+    }
+
+    ChunkedInputStream(BufferedByteSource src, byte[] preBuf, int preBufOff, int preBufLen) {
+        this(src, preBuf, preBufOff, preBufLen, new MutableHeaderMap());
+    }
+
+    @Override
+    public boolean fullyRead() {
+        return done;
     }
 
     @Override
@@ -29,7 +54,7 @@ final class ChunkedInputStream extends InputStream {
             if (chunkRemaining == 0) { consumeTrailers(); done = true; return -1; }
         }
         int b = src.read();
-        if (b >= 0 && --chunkRemaining == 0) { src.read(); src.read(); } // consume trailing \r\n
+        if (b >= 0 && --chunkRemaining == 0) consumeChunkTerminator();
         return b;
     }
 
@@ -43,30 +68,154 @@ final class ChunkedInputStream extends InputStream {
         int n = src.read(buf, off, Math.min(len, chunkRemaining));
         if (n > 0) {
             chunkRemaining -= n;
-            if (chunkRemaining == 0) { src.read(); src.read(); } // consume trailing \r\n
+            if (chunkRemaining == 0) consumeChunkTerminator();
         }
         return n;
     }
 
+    /** Validates and consumes the CRLF that terminates every chunk's data (RFC 9112 §7.1.1). */
+    private void consumeChunkTerminator() throws IOException {
+        int cr = src.read();
+        int lf = src.read();
+        if (cr != '\r' || lf != '\n') {
+            throw new MalformedRequestException(400, "Malformed chunk terminator");
+        }
+    }
+
+    /**
+     * Reads one chunk-size line: hex digits, an optional {@code ;}-prefixed chunk-extension
+     * (discarded — RFC 9112 §7.1.1 permits ignoring extensions this server does not recognise),
+     * then CRLF. Bounded per {@code Http1Limits} against: more than 16 hex digits (a chunk size
+     * cannot legitimately need more — {@code Long.MAX_VALUE} is 16 hex digits), a size above
+     * {@link Http1Limits#MAX_CHUNK_SIZE}, an extension longer than
+     * {@link Http1Limits#MAX_CHUNK_EXT_LENGTH}, and more than
+     * {@link Http1Limits#MAX_CHUNKS_PER_BODY} chunks per body — all defences against a peer
+     * that is technically well-formed but deliberately expensive to parse.
+     */
     private int readChunkSize() throws IOException {
+        if (++chunksSeen > Http1Limits.MAX_CHUNKS_PER_BODY) {
+            throw new MalformedRequestException(413, "Too many chunks");
+        }
+
         long size = 0;
-        int  b;
-        while ((b = src.read()) != -1) {
-            if      (b >= '0' && b <= '9') size = (size << 4) | (b - '0');
-            else if (b >= 'a' && b <= 'f') size = (size << 4) | (b - 'a' + 10);
-            else if (b >= 'A' && b <= 'F') size = (size << 4) | (b - 'A' + 10);
-            else { while ((b = src.read()) != -1 && b != '\n'); break; } // ext or \r\n
-            if (size > Integer.MAX_VALUE) throw new IOException("Chunk size exceeds 2 GB limit");
+        int  digits = 0;
+        int  b = src.read();
+        while (isHexDigit(b)) {
+            if (++digits > 16) throw new MalformedRequestException(400, "Chunk size line too long");
+            size = (size << 4) | hexValue(b);
+            if (size > Http1Limits.MAX_CHUNK_SIZE) {
+                throw new MalformedRequestException(413, "Chunk size exceeds configured maximum");
+            }
+            b = src.read();
+        }
+        if (digits == 0) throw new MalformedRequestException(400, "Malformed chunk size");
+
+        int extLen = 0;
+        while (b != -1 && b != '\r') {
+            if (++extLen > Http1Limits.MAX_CHUNK_EXT_LENGTH) {
+                throw new MalformedRequestException(400, "Chunk extension too long");
+            }
+            b = src.read();
+        }
+        if (b != '\r' || src.read() != '\n') {
+            throw new MalformedRequestException(400, "Malformed chunk size line terminator");
         }
         return (int) size;
     }
 
-    // Reads and discards trailer headers until the empty line that terminates the chunked body.
+    private static boolean isHexDigit(int b) {
+        return (b >= '0' && b <= '9') || (b >= 'a' && b <= 'f') || (b >= 'A' && b <= 'F');
+    }
+
+    private static int hexValue(int b) {
+        if (b <= '9') return b - '0';
+        if (b <= 'F') return b - 'A' + 10;
+        return b - 'a' + 10;
+    }
+
+    /**
+     * Reads and discards trailer headers until the empty line that terminates the chunked body
+     * (RFC 9112 §7.1.2). Bounded by {@link Http1Limits#MAX_TRAILER_COUNT} and
+     * {@link Http1Limits#MAX_HEADER_VALUE_LENGTH} — without a bound, a peer could follow the
+     * final chunk with an unbounded trailer section purely to waste CPU discarding it. Trailers
+     * ({@code Request.trailers()}).
+     */
     private void consumeTrailers() throws IOException {
+        int trailerCount = 0;
         while (true) {
             int b = src.read();
-            if (b == -1 || b == '\r') { src.read(); return; } // empty line — done
-            while ((b = src.read()) != -1 && b != '\n');      // skip non-empty trailer line
+            if (b == -1) throw new MalformedRequestException(400, "Truncated trailer section");
+            if (b == '\r') {
+                if (src.read() != '\n') {
+                    throw new MalformedRequestException(400, "Malformed trailer section terminator");
+                }
+                return; // empty line — trailer section done
+            }
+            if (++trailerCount > Http1Limits.MAX_TRAILER_COUNT) {
+                throw new MalformedRequestException(431, "Too many trailers");
+            }
+            int lineLen = 0;
+            trailerLine[lineLen++] = (byte) b;
+            while ((b = src.read()) != -1 && b != '\n') {
+                if (lineLen == trailerLine.length) {
+                    throw new MalformedRequestException(431, "Trailer line too long");
+                }
+                trailerLine[lineLen++] = (byte) b;
+            }
+            if (b != '\n' || lineLen == 0 || trailerLine[lineLen - 1] != '\r') {
+                throw new MalformedRequestException(400, "Malformed trailer line");
+            }
+            addTrailer(lineLen - 1);
         }
     }
+
+    private void addTrailer(int lineLength) throws MalformedRequestException {
+        int colon = -1;
+        for (int i = 0; i < lineLength; i++) {
+            if (trailerLine[i] == ':') { colon = i; break; }
+        }
+        if (colon <= 0) throw new MalformedRequestException(400, "Malformed trailer field");
+        for (int i = 0; i < colon; i++) {
+            int c = trailerLine[i] & 0xff;
+            if (!isToken(c)) {
+                throw new MalformedRequestException(400, "Invalid trailer field name");
+            }
+        }
+        int valueStart = colon + 1;
+        while (valueStart < lineLength
+            && (trailerLine[valueStart] == ' ' || trailerLine[valueStart] == '\t')) valueStart++;
+        int valueEnd = lineLength;
+        while (valueEnd > valueStart
+            && (trailerLine[valueEnd - 1] == ' ' || trailerLine[valueEnd - 1] == '\t')) valueEnd--;
+        if (forbidden(trailerLine, colon)) {
+            throw new MalformedRequestException(400, "Forbidden trailer field");
+        }
+        trailers.add(trailerLine, 0, colon, trailerLine, valueStart, valueEnd - valueStart);
+    }
+
+    private static boolean isToken(int c) {
+        return (c >= '0' && c <= '9')
+            || (c >= 'A' && c <= 'Z')
+            || (c >= 'a' && c <= 'z')
+            || c == '!' || c == '#' || c == '$' || c == '%' || c == '&' || c == '\''
+            || c == '*' || c == '+' || c == '-' || c == '.' || c == '^' || c == '_'
+            || c == '`' || c == '|' || c == '~';
+    }
+
+    private static boolean forbidden(byte[] name, int length) {
+        return asciiEquals(name, length, "content-length")
+            || asciiEquals(name, length, "transfer-encoding")
+            || asciiEquals(name, length, "host")
+            || asciiEquals(name, length, "trailer");
+    }
+
+    private static boolean asciiEquals(byte[] bytes, int length, String expected) {
+        if (length != expected.length()) return false;
+        for (int i = 0; i < length; i++) {
+            int c = bytes[i] & 0xff;
+            if (c >= 'A' && c <= 'Z') c += 32;
+            if (c != expected.charAt(i)) return false;
+        }
+        return true;
+    }
 }

flash/src/main/java/dev/relism/flash/HttpServer.java

@@ -1,564 +0,0 @@
-package dev.relism.flash;
-
-import dev.relism.flash.extension.FlashApp;
-import dev.relism.flash.extension.FlashConfiguration;
-import dev.relism.flash.http.ContentType;
-import dev.relism.flash.http.HttpMethod;
-import dev.relism.flash.http.HttpStatus;
-import dev.relism.flash.models.Request;
-import dev.relism.flash.models.RequestHandler;
-import dev.relism.flash.models.Response;
-import dev.relism.flash.routing.AbstractRouter;
-import dev.relism.flash.routing.AbstractWsRouter;
-import dev.relism.flash.tls.TlsConfig;
-import dev.relism.flash.websocket.WebSocketFrame;
-import dev.relism.flash.websocket.WebSocketHandler;
-import dev.relism.flash.websocket.WebSocketSession;
-import dev.relism.fpr.core.ByteView;
-
-import lombok.extern.slf4j.Slf4j;
-
-import javax.net.ssl.SSLServerSocket;
-import javax.net.ssl.SSLSocket;
-
-import java.io.*;
-import java.net.InetSocketAddress;
-import java.net.ServerSocket;
-import java.net.Socket;
-import java.nio.charset.StandardCharsets;
-import java.security.MessageDigest;
-import java.security.NoSuchAlgorithmException;
-import java.util.ArrayList;
-import java.util.Base64;
-import java.util.List;
-import java.util.Set;
-import java.util.concurrent.*;
-import java.util.concurrent.atomic.AtomicInteger;
-import java.util.concurrent.atomic.AtomicReference;
-
-/**
- * Pure I/O transport layer. Owns one {@link ServerSocket} per configured listener (plain or
- * TLS), the virtual-thread executor, and the keep-alive accept loop. Routing is delegated to
- * HTTP and WS routers — identically, regardless of which listener accepted the connection.
- *
- * <p>TLS is a transport-level concern only: once a {@link BoundListener} is bound, an accepted
- * {@link Socket} is either plain or an {@code SSLSocket} indistinguishably from here on —
- * {@link #process} never branches on it. This is also why WSS needs no separate code path from
- * WS: the WebSocket upgrade happens over whatever transport {@link #process} was handed.
- *
- * <h3>Allocation model</h3>
- * <ul>
- *   <li>{@code LONG_BUF} (20 bytes) and {@code STREAM_RELAY_BUFFER} (8 KB, for a streaming
- *       {@link Response} body — see {@link #writeStreamingBody}) are the only {@link ThreadLocal}s
- *       kept here. Both are per-connection, not per-request: one virtual thread runs a
- *       connection's whole keep-alive request loop (see {@link #process}), so a handler that
- *       streams a large response on every request allocates its relay buffer once per
- *       connection, not once per request.</li>
- *   <li>WS handshake SHA-1: {@link ThreadLocal}&lt;{@link MessageDigest}&gt; — one per
- *       accept thread (there are now {@code ACCEPT_THREADS} of them, not one).</li>
- * </ul>
- */
-@Slf4j
-class HttpServer implements ServerHandle {
-
-    // ── Tuning constants ──────────────────────────────────────────────────────
-
-    /**
-     * Number of platform threads competing on {@code serverSocket.accept()}.
-     * Rule of thumb: number of available CPU cores, capped at 8.
-     * More than this rarely helps — accept is cheap; the bottleneck is usually
-     * the virtual-thread executor dispatching the connection handler.
-     */
-    private static final int ACCEPT_THREADS  = Math.min(Runtime.getRuntime().availableProcessors(), 8);
-
-    /**
-     * TCP listen backlog. The kernel holds up to this many fully-established
-     * (SYN+ACK sent, ACK received) connections waiting for accept().
-     * 4096 is safe on Linux; /proc/sys/net/core/somaxconn must be >= this value,
-     * or the kernel silently caps it. Raise somaxconn if needed:
-     *   sysctl -w net.core.somaxconn=4096
-     */
-    private static final int ACCEPT_BACKLOG  = 4096;
-
-    /**
-     * Socket send/receive buffer sizes. Matched to the WS frame read buffer
-     * ({@link FlashConfiguration#getWsFrameBufferSize()}) so the kernel never
-     * needs to fragment a full frame into multiple TCP segments on the receive
-     * side, and never blocks a write waiting for the send buffer to drain.
-     *
-     * Linux default is 87380 bytes (rmem) / 16384 bytes (wmem). We raise both
-     * to 256 KB — a good fit for up to ~250 KB WS frames with no partial reads.
-     */
-    private static final int SOCKET_BUF_SIZE = 256 * 1024;
-
-    // ── Instance fields ───────────────────────────────────────────────────────
-
-    private final FlashConfiguration  configuration;
-    private final List<BoundListener> boundListeners;
-    private final AbstractRouter      router;
-    private final AbstractWsRouter    wsRouter;
-    private final ExecutorService     executorService  = Executors.newVirtualThreadPerTaskExecutor();
-    private final Set<Socket>         activeSockets    = ConcurrentHashMap.newKeySet();
-    private volatile boolean          stopped          = false;
-
-    /** Latch that reaches 0 when all accept threads, across all listeners, have exited. */
-    private final CountDownLatch      acceptLatch;
-
-    /** One bound listener socket (plain or TLS) plus whether it is TLS, for logging only. */
-    private record BoundListener(ServerSocket socket, boolean secure) {}
-
-    // ── Static byte constants (written once, read-only on hot path) ──────────
-
-    private static final byte[] HTTP_1_1              = "HTTP/1.1 ".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] CRLF                  = "\r\n".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] CONTENT_TYPE          = "Content-Type: ".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] CONTENT_LENGTH        = "Content-Length: ".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] CONNECTION_CLOSE      = "Connection: close\r\n".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] CONNECTION_KEEPALIVE  = "Connection: keep-alive\r\n".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] TRANSFER_CHUNKED      = "Transfer-Encoding: chunked\r\n".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] FINAL_CHUNK           = "0\r\n\r\n".getBytes(StandardCharsets.UTF_8);
-    private static final byte[] UNKNOWN_STATUS_SUFFIX = " Unknown".getBytes(StandardCharsets.UTF_8);
-
-    private static final byte[] WS_HANDSHAKE_PREFIX =
-            ("HTTP/1.1 101 Switching Protocols\r\n" +
-                    "Upgrade: websocket\r\n" +
-                    "Connection: Upgrade\r\n" +
-                    "Sec-WebSocket-Accept: ")
-                    .getBytes(StandardCharsets.ISO_8859_1);
-    private static final byte[] WS_HANDSHAKE_SUFFIX =
-            "\r\n\r\n".getBytes(StandardCharsets.ISO_8859_1);
-    private static final byte[] WS_REJECT_400 =
-            "HTTP/1.1 400 Bad Request\r\nContent-Length: 0\r\nConnection: close\r\n\r\n"
-                    .getBytes(StandardCharsets.ISO_8859_1);
-
-    private static final byte[] WS_GUID_BYTES =
-            "258EAFA5-E914-47DA-95CA-C5AB0DC85B11".getBytes(StandardCharsets.ISO_8859_1);
-
-    private static final ThreadLocal<MessageDigest> SHA1 =
-            ThreadLocal.withInitial(() -> {
-                try { return MessageDigest.getInstance("SHA-1"); }
-                catch (NoSuchAlgorithmException e) { throw new RuntimeException(e); }
-            });
-
-    private static final ThreadLocal<byte[]> LONG_BUF = ThreadLocal.withInitial(() -> new byte[20]);
-
-    /**
-     * Relay buffer for copying a streaming {@link Response} body to the client — shared by
-     * {@link #writeStreamingBody}'s non-chunked path and {@link #writeChunked}, so both draw
-     * from the same reused array instead of each allocating its own {@code byte[8192]} (the
-     * non-chunked path previously relied on {@link InputStream#transferTo}, which allocates
-     * internally on every call). Sized to match the pre-existing behavior this replaces, not
-     * newly tuned — not exposed as a {@link FlashConfiguration} tunable since nothing here
-     * needed one before.
-     */
-    private static final int STREAM_RELAY_BUFFER_SIZE = 8192;
-    private static final ThreadLocal<byte[]> STREAM_RELAY_BUFFER =
-            ThreadLocal.withInitial(() -> new byte[STREAM_RELAY_BUFFER_SIZE]);
-
-    private static final int SHA1_LEN      = 20;
-    private static final int WS_ACCEPT_LEN = 28;
-
-    // ── Constructor ───────────────────────────────────────────────────────────
-
-    HttpServer(FlashConfiguration configuration, AbstractRouter router, AbstractWsRouter wsRouter) throws IOException {
-        this.configuration = configuration;
-        this.router        = router;
-        this.wsRouter      = wsRouter;
-
-        List<FlashConfiguration.Listener> specs = configuration.getListeners().isEmpty()
-                ? List.of(new FlashConfiguration.Listener(
-                        configuration.getPort(), configuration.getHost(), configuration.getTls()))
-                : configuration.getListeners();
-
-        List<BoundListener> bound = new ArrayList<>(specs.size());
-        for (FlashConfiguration.Listener spec : specs) bound.add(bind(spec));
-        this.boundListeners = List.copyOf(bound);
-        this.acceptLatch    = new CountDownLatch(ACCEPT_THREADS * boundListeners.size());
-
-        for (BoundListener bl : boundListeners) {
-            log.info("HttpServer bound on {}:{} (tls={}, backlog={}, acceptThreads={})",
-                    bl.socket().getInetAddress(), bl.socket().getLocalPort(), bl.secure(),
-                    ACCEPT_BACKLOG, ACCEPT_THREADS);
-        }
-    }
-
-    /**
-     * Binds one listener. A TLS listener gets its {@link ServerSocket} from
-     * {@link TlsConfig#serverSocketFactory()} instead of {@code new ServerSocket()}, and its
-     * protocol/client-auth parameters from {@link TlsConfig#applyTo} — reuse-address, receive
-     * buffer size, backlog and the bind call itself are identical either way. TLS only changes
-     * which bytes come out of {@code accept()}; it never changes how the accept loop, or
-     * anything downstream of it, treats them.
-     */
-    private static BoundListener bind(FlashConfiguration.Listener spec) throws IOException {
-        TlsConfig tls = spec.tls();
-
-        ServerSocket socket = tls != null ? tls.serverSocketFactory().createServerSocket() : new ServerSocket();
-        // setReuseAddress(true) must be called BEFORE bind().
-        socket.setReuseAddress(true);
-        socket.setReceiveBufferSize(SOCKET_BUF_SIZE);
-        if (tls != null) tls.applyTo((SSLServerSocket) socket);
-
-        InetSocketAddress addr = spec.host() != null
-                ? new InetSocketAddress(spec.host(), spec.port())
-                : new InetSocketAddress(spec.port());
-        socket.bind(addr, ACCEPT_BACKLOG);
-
-        return new BoundListener(socket, tls != null);
-    }
-
-    // ── Lifecycle ─────────────────────────────────────────────────────────────
-
-    @Override
-    public void start() {
-        for (int li = 0; li < boundListeners.size(); li++) {
-            BoundListener listener = boundListeners.get(li);
-            for (int i = 0; i < ACCEPT_THREADS; i++) {
-                Thread.ofPlatform()
-                        .name("flash-accept-" + li + "-" + i)
-                        .daemon(false)
-                        .start(() -> acceptLoop(listener));
-            }
-        }
-    }
-
-    @Override
-    public void startAndBlock() {
-        start();
-        try { acceptLatch.await(); }
-        catch (InterruptedException e) { Thread.currentThread().interrupt(); }
-    }
-
-    /**
-     * Single accept loop body — runs on each of the {@code ACCEPT_THREADS}
-     * platform threads bound to one {@code listener}. All threads for that listener block on
-     * the same {@link ServerSocket}; the JVM ensures only one wakes per incoming connection
-     * (no thundering herd). Other listeners' accept threads are entirely independent.
-     */
-    private void acceptLoop(BoundListener listener) {
-        try {
-            while (!stopped) {
-                try {
-                    process(listener.socket().accept());
-                } catch (IOException e) {
-                    if (!stopped) log.error("Accept error", e);
-                }
-            }
-        } finally {
-            acceptLatch.countDown();
-        }
-    }
-
-    @Override
-    public CompletableFuture<Void> stop() {
-        return CompletableFuture.runAsync(() -> {
-            stopped = true;
-            for (BoundListener bl : boundListeners) {
-                try { bl.socket().close(); } catch (IOException e) { log.error("Error closing server socket", e); }
-            }
-            activeSockets.forEach(s -> { try { s.close(); } catch (IOException ignored) {} });
-            executorService.shutdown();
-            try {
-                if (!executorService.awaitTermination(30, TimeUnit.SECONDS))
-                    executorService.shutdownNow();
-            } catch (InterruptedException e) {
-                executorService.shutdownNow();
-                Thread.currentThread().interrupt();
-            }
-        });
-    }
-
-    // ── Hot-path ──────────────────────────────────────────────────────────────
-
-    private void process(Socket socket) {
-        try {
-            executorService.submit(() -> {
-                activeSockets.add(socket);
-                try (socket;
-                     InputStream  in  = socket.getInputStream();
-                     OutputStream out = new BufferedOutputStream(socket.getOutputStream())) {
-
-                    // TCP_NODELAY: disable Nagle's algorithm.
-                    // Small WS frames (< MSS) are sent immediately rather than
-                    // waiting up to 200 ms for more data to coalesce. Latency
-                    // drops significantly at the cost of slightly more TCP segments
-                    // under sustained bulk transfer — acceptable for interactive WS.
-                    socket.setTcpNoDelay(true);
-                    socket.setSendBufferSize(SOCKET_BUF_SIZE);
-
-                    // rawOut is the unbuffered socket stream — passed to WebSocketSession
-                    // directly. WS writes are already bulk (header + payload in two calls);
-                    // with TCP_NODELAY the kernel ships them without Nagle delay, so no
-                    // userspace buffer is needed and no flush() is required per frame.
-                    // HTTP responses continue to use the BufferedOutputStream (out) because
-                    // writeResponse() does many small individual writes that benefit from
-                    // userspace coalescing before a single syscall.
-                    OutputStream rawOut = socket.getOutputStream();
-
-                    RequestParser parser = new RequestParser(
-                            configuration.getMaxHeaderBufferSize(),
-                            (InetSocketAddress) socket.getRemoteSocketAddress(),
-                            socket instanceof SSLSocket sslSocket ? sslSocket : null);
-
-                    while (!stopped) {
-                        Request request = parser.parse(in);
-                        if (request == null) break;
-
-                        if (request.method() == HttpMethod.GET && isWebSocketUpgrade(request)) {
-                            WebSocketHandler wsHandler = wsRouter.route(request);
-                            if (wsHandler == null) {
-                                out.write(WS_REJECT_400);
-                                out.flush();
-                                break;
-                            }
-                            // Flush buffered HTTP bytes (the 101 response) before WebSocketSession
-                            // takes over rawOut — otherwise the handshake reply stays stuck in
-                            // the BufferedOutputStream buffer and the client never sees it.
-                            performHandshake(out, request);
-                            out.flush();
-                            request.drain();
-                            WebSocketSession session = new WebSocketSession(
-                                    in, rawOut, configuration.getWsFrameBufferSize(), request, false);
-                            runWsLoop(session, wsHandler);
-                            return;
-                        }
-
-                        boolean keepAlive = isKeepAlive(request);
-                        Response response = new Response(200, ContentType.TEXT_PLAIN);
-
-                        RequestHandler handler = router.route(request);
-                        if (handler == null) handler = router.getNotFoundHandler();
-
-                        try {
-                            Object result = handler.handle(request, response);
-                            if (result instanceof Response r) response = r;
-                            else if (result != null) response.setBody(result);
-                        } catch (Exception ex) {
-                            Object result = router.getExceptionHandler().handle(ex, request, response);
-                            if (result instanceof Response r) response = r;
-                            else if (result != null) response.setBody(result);
-                        }
-
-                        writeResponse(out, response, keepAlive);
-                        request.drain();
-                        if (!keepAlive) break;
-                    }
-
-                } catch (IOException e) {
-                    if (!stopped) {
-                        if (e instanceof java.net.SocketException)
-                            log.debug("Connection closed: {}", e.getMessage());
-                        else
-                            log.error("I/O error handling request", e);
-                    }
-                } catch (Exception e) {
-                    // Anything not an IOException here means a collaborator misbehaved on the TLS
-                    // handshake path — most likely a custom TlsConfig#ofContext KeyManager/
-                    // TrustManager throwing (e.g. a failed DB lookup or on-demand cert issuance).
-                    // That failure is isolated to this one virtual thread/connection: the
-                    // try-with-resources above still closes the socket, the finally below still
-                    // runs, and the accept loop (a different thread entirely) never sees this.
-                    if (!stopped) log.error("Unexpected error handling connection", e);
-                } finally {
-                    activeSockets.remove(socket);
-                }
-            });
-        } catch (RejectedExecutionException ignored) {
-            try { socket.close(); } catch (IOException e) { log.debug("Error closing socket on shutdown", e); }
-        }
-    }
-
-    // ── WebSocket upgrade detection (zero-alloc) ──────────────────────────────
-
-    private static boolean isWebSocketUpgrade(Request request) {
-        ByteView upgrade = request.getRequestLine().getHeaders().view("Upgrade");
-        if (upgrade == null) return false;
-        if (!tokenEqualsIgnoreCase(upgrade, 0, upgrade.length(), "websocket")) return false;
-        return connectionContainsUpgrade(request);
-    }
-
-    private static boolean connectionContainsUpgrade(Request request) {
-        ByteView conn = request.getRequestLine().getHeaders().view("Connection");
-        if (conn == null) return false;
-        int len = conn.length(), i = 0;
-        while (i < len) {
-            while (i < len && conn.byteAt(i) == ' ') i++;
-            int start = i;
-            while (i < len && conn.byteAt(i) != ',') i++;
-            if (tokenEqualsIgnoreCase(conn, start, i, "upgrade")) return true;
-            i++;
-        }
-        return false;
-    }
-
-    private static boolean tokenEqualsIgnoreCase(ByteView view, int start, int end, String token) {
-        int tlen = token.length();
-        int wlen = end - start;
-        while (wlen > 0 && view.byteAt(start + wlen - 1) == ' ') wlen--;
-        if (wlen != tlen) return false;
-        for (int i = 0; i < tlen; i++) {
-            byte b = view.byteAt(start + i);
-            if (b >= 'A' && b <= 'Z') b += 32;
-            if (b != (byte) token.charAt(i)) return false;
-        }
-        return true;
-    }
-
-    // ── WebSocket handshake ───────────────────────────────────────────────────
-
-    private void performHandshake(OutputStream out, Request request) throws IOException {
-        ByteView keyView = request.getRequestLine().getHeaders().view("Sec-WebSocket-Key");
-        if (keyView == null) throw new IOException("Missing Sec-WebSocket-Key header");
-
-        MessageDigest sha1 = SHA1.get();
-        sha1.reset();
-        for (int i = 0, len = keyView.length(); i < len; i++) sha1.update(keyView.byteAt(i));
-        sha1.update(WS_GUID_BYTES);
-
-        byte[] accept = Base64.getEncoder().encode(sha1.digest());
-
-        out.write(WS_HANDSHAKE_PREFIX);
-        out.write(accept);
-        out.write(WS_HANDSHAKE_SUFFIX);
-        out.flush();
-    }
-
-    // ── WebSocket session loop ────────────────────────────────────────────────
-
-    private void runWsLoop(WebSocketSession session, WebSocketHandler handler) {
-        handler.onOpen(session);
-        WebSocketFrame frame = new WebSocketFrame();
-        try {
-            while (session.isOpen()) {
-                if (!session.readFrame(frame)) break;
-                switch (frame.opcode()) {
-                    case WebSocketFrame.OP_TEXT, WebSocketFrame.OP_BINARY
-                            -> handler.onMessage(session, frame);
-                    case WebSocketFrame.OP_CLOSE
-                            -> session.closeFromPeer(frame);
-                    case WebSocketFrame.OP_PING
-                            -> session.sendPong(frame);
-                    case WebSocketFrame.OP_PONG -> { /* heartbeat ack, no-op */ }
-                }
-            }
-        } catch (IOException e) {
-            handler.onError(session, e);
-        } finally {
-            handler.onClose(session, session.closeCode());
-            session.forceClose();
-        }
-    }
-
-    // ── HTTP keep-alive detection ─────────────────────────────────────────────
-
-    private static boolean isKeepAlive(Request request) {
-        if (request.headerEquals("Connection", "close")) return false;
-        ByteView protocol = request.getRequestLine().getProtocol();
-        int plen = protocol.length();
-        if (plen == 8) {
-            byte minor = protocol.byteAt(7);
-            if (minor == '1') return true;
-            if (minor == '0') return request.headerEquals("Connection", "keep-alive");
-        }
-        log.debug("Unrecognised protocol '{}', treating as close", protocol);
-        return false;
-    }
-
-    // ── Response serialisation ────────────────────────────────────────────────
-
-    private static void writeResponse(OutputStream out, Response response, boolean keepAlive) throws IOException {
-        out.write(HTTP_1_1);
-        byte[] statusBytes = response.getStatusBytes();
-        if (statusBytes != null) out.write(statusBytes);
-        else writeStatusPhrase(out, response.getStatusCode());
-        out.write(CRLF);
-        out.write(CONTENT_TYPE);
-        out.write(response.getContentType());
-        out.write(CRLF);
-        response.writeHeaders(out);
-
-        if (response.isStreaming()) {
-            writeStreamingBody(out, response, keepAlive);
-        } else {
-            byte[] body = response.getBody();
-            out.write(CONTENT_LENGTH);
-            writeLong(out, body != null ? body.length : 0);
-            out.write(CRLF);
-            out.write(keepAlive ? CONNECTION_KEEPALIVE : CONNECTION_CLOSE);
-            out.write(CRLF);
-            if (body != null) out.write(body);
-        }
-        out.flush();
-    }
-
-    private static void writeStreamingBody(OutputStream out, Response response, boolean keepAlive) throws IOException {
-        if (!response.isChunked()) {
-            out.write(CONTENT_LENGTH);
-            writeLong(out, response.getStreamLength());
-            out.write(CRLF);
-            out.write(keepAlive ? CONNECTION_KEEPALIVE : CONNECTION_CLOSE);
-            out.write(CRLF);
-            relay(response.getStream(), out);
-        } else {
-            out.write(TRANSFER_CHUNKED);
-            out.write(keepAlive ? CONNECTION_KEEPALIVE : CONNECTION_CLOSE);
-            out.write(CRLF);
-            writeChunked(out, response.getStream());
-        }
-    }
-
-    /**
-     * Copies {@code in} to {@code out} until EOF, same contract as {@link InputStream#transferTo}
-     * — but via {@link #STREAM_RELAY_BUFFER} instead of a fresh {@code byte[]} per call, which is
-     * what {@code transferTo}'s own (JDK-internal) implementation would otherwise allocate on
-     * every streamed response.
-     */
-    private static void relay(InputStream in, OutputStream out) throws IOException {
-        byte[] buf = STREAM_RELAY_BUFFER.get();
-        int n;
-        while ((n = in.read(buf)) > 0) out.write(buf, 0, n);
-    }
-
-    private static void writeStatusPhrase(OutputStream out, int statusCode) throws IOException {
-        byte[] phrase = HttpStatus.bytesForCode(statusCode);
-        if (phrase != null) out.write(phrase);
-        else { writeLong(out, statusCode); out.write(UNKNOWN_STATUS_SUFFIX); }
-    }
-
-    private static void writeLong(OutputStream out, long value) throws IOException {
-        if (value == 0) { out.write('0'); return; }
-        byte[] buf = LONG_BUF.get();
-        int pos = buf.length;
-        boolean neg = value < 0;
-        if (neg) value = -value;
-        do { buf[--pos] = (byte) ('0' + value % 10); value /= 10; } while (value > 0);
-        if (neg) buf[--pos] = '-';
-        out.write(buf, pos, buf.length - pos);
-    }
-
-    private static void writeChunked(OutputStream out, InputStream stream) throws IOException {
-        byte[] buf = STREAM_RELAY_BUFFER.get();
-        int n;
-        while ((n = stream.read(buf)) > 0) {
-            writeHex(out, n);
-            out.write(CRLF);
-            out.write(buf, 0, n);
-            out.write(CRLF);
-        }
-        out.write(FINAL_CHUNK);
-    }
-
-    private static void writeHex(OutputStream out, int value) throws IOException {
-        int shift   = 28;
-        boolean leading = true;
-        while (shift >= 0) {
-            int digit = (value >>> shift) & 0xF;
-            if (digit != 0 || !leading) {
-                leading = false;
-                out.write(digit < 10 ? '0' + digit : 'a' + digit - 10);
-            }
-            shift -= 4;
-        }
-        if (leading) out.write('0');
-    }
-}

flash/src/main/java/dev/relism/flash/RequestParser.java

@@ -1,17 +1,22 @@
 package dev.relism.flash;
 
+import dev.relism.flash.bytes.ByteScan;
+import dev.relism.flash.exceptions.MalformedRequestException;
+import dev.relism.flash.http.Http1Limits;
 import dev.relism.flash.http.HttpMethod;
-import dev.relism.flash.models.HeaderMap;
+import dev.relism.flash.models.Http1HeaderMap;
+import dev.relism.flash.models.MutableHeaderMap;
 import dev.relism.flash.models.Request;
+import dev.relism.flash.models.RequestBody;
 import dev.relism.flash.models.RequestLine;
 import dev.relism.flash.routing.routers.fastpathrouter.FastPathViews;
+import dev.relism.flash.transport.BufferedByteSource;
 
 import lombok.extern.slf4j.Slf4j;
 
 import javax.net.ssl.SSLSocket;
 
 import java.io.IOException;
-import java.io.InputStream;
 import java.net.InetSocketAddress;
 import java.util.Arrays;
 
@@ -35,6 +40,15 @@ import java.util.Arrays;
  * to the <em>next</em> request. They are snapshotted at the top of {@link #parse}
  * and reset to {@code 0/0} before any work begins, so an exception thrown mid-parse
  * leaves the fields clean rather than pointing at stale data from a previous request.
+ *
+ * Anything wrong with the request itself — smuggling-relevant ambiguity, an over-limit
+ * header, a malformed byte where the grammar forbids one — is reported as a
+ * {@link MalformedRequestException} carrying the exact status the caller must respond with.
+ * This is distinct from {@link IOException}, which still means "the socket failed" (EOF,
+ * reset, timeout). The caller ({@code HttpServer.process}) must always close the connection
+ * after a {@link MalformedRequestException}, never keep it alive — RFC 9112 §6.1's rationale
+ * for rejecting {@code Content-Length} + {@code Transfer-Encoding} outright is exactly that a
+ * kept-alive connection after a disputed request boundary is what a smuggling attack needs.
  */
 @Slf4j
 public class RequestParser {
@@ -43,7 +57,17 @@ public class RequestParser {
     private final int               maxHeaderBufferSize;
     private final InetSocketAddress remoteAddress;
     private final SSLSocket         sslSocket;
-    private final HeaderMap         headerMap = new HeaderMap();
+    private final Http1HeaderMap    headerMap   = new Http1HeaderMap();
+    private final MutableHeaderMap  trailerMap  = new MutableHeaderMap();
+    // request — same idiom as headerMap above.
+    private final RequestLine       requestLine = new RequestLine();
+    private final Request           request     = new Request();
+    private final RequestBody       requestBody = new RequestBody();
+    // RequestLine/Response themselves. queryView is only reset and used when a query string is
+    // actually present; RequestLine.getQuery() must keep returning null otherwise (see reset()).
+    private final FastPathViews.RequestByteView pathView     = new FastPathViews.RequestByteView(null, 0, 0);
+    private final FastPathViews.RequestByteView queryView    = new FastPathViews.RequestByteView(null, 0, 0);
+    private final FastPathViews.RequestByteView protocolView = new FastPathViews.RequestByteView(null, 0, 0);
     private byte[] buffer;
 
     // Unconsumed bytes belonging to the NEXT request.
@@ -65,6 +89,19 @@ public class RequestParser {
         this.buffer = new byte[Math.min(INITIAL_BUFFER_SIZE, maxHeaderBufferSize)];
     }
 
+    /**
+     * Whether bytes from a previous {@link #parse} call are already buffered and ready to be
+     * consumed by the next call without reading anything further from the source — the HTTP
+     * pipelining case. The caller (the connection loop) uses this to decide whether it is safe
+     * to skip waiting for "the next request has started arriving": if bytes are already
+     * buffered, the next request has, by definition, already started (and may even be
+     * complete), so an idle-timeout wait on the underlying source would wait for bytes that
+     * were never going to arrive there — they are already here.
+     */
+    public boolean hasBufferedBytes() {
+        return bufLen > 0;
+    }
+
     /**
      * Parses the next HTTP request from {@code in}.
      *
@@ -75,9 +112,12 @@ public class RequestParser {
      * the same parser instance is reused after an error.
      *
      * @return the parsed {@link Request}, or {@code null} on clean EOF.
-     * @throws IOException on malformed headers or I/O failure.
+     * @throws MalformedRequestException if the request violates the HTTP/1.1 grammar or a
+     *         configured safety limit — carries the exact status to respond with.
+     * @throws IOException on genuine I/O failure (socket reset, timeout).
      */
-    public Request parse(InputStream in) throws IOException {
+    public Request parse(BufferedByteSource in) throws IOException {
+        trailerMap.reset();
         // Snapshot leftover bytes from the previous request, then reset immediately.
         // Any exception thrown below leaves bufBase/bufLen at 0 — safe state.
         int base      = bufBase;
@@ -85,7 +125,7 @@ public class RequestParser {
         bufBase = 0;
         bufLen  = 0;
 
-        int headerEndIdx = totalRead > 0 ? findEndOfHeader(buffer, base, base + totalRead) : -1;
+        int headerEndIdx = totalRead > 0 ? ByteScan.indexOfCrLfCrLf(buffer, base, base + totalRead) : -1;
         while (headerEndIdx == -1) {
             if (base + totalRead == buffer.length) {
                 if (base > 0) {
@@ -94,7 +134,8 @@ public class RequestParser {
                     System.arraycopy(buffer, base, buffer, 0, totalRead);
                     base = 0;
                 } else if (buffer.length >= maxHeaderBufferSize) {
-                    throw new IOException("Request headers exceed " + maxHeaderBufferSize + " bytes");
+                    throw new MalformedRequestException(431,
+                            "Request headers exceed " + maxHeaderBufferSize + " bytes");
                 } else {
                     buffer = Arrays.copyOf(buffer, Math.min(buffer.length * 2, maxHeaderBufferSize));
                 }
@@ -103,65 +144,133 @@ public class RequestParser {
             if (n <= 0) break;
             int prevTotal = totalRead;
             totalRead += n;
-            headerEndIdx = findEndOfHeader(buffer, base + Math.max(0, prevTotal - 3), base + totalRead);
+            headerEndIdx = ByteScan.indexOfCrLfCrLf(buffer, base + Math.max(0, prevTotal - 3), base + totalRead);
         }
         if (totalRead <= 0) return null;
         if (headerEndIdx == -1) {
-            throw new IOException("Request headers exceed " + maxHeaderBufferSize + " bytes");
+            throw new MalformedRequestException(431,
+                    "Request headers exceed " + maxHeaderBufferSize + " bytes");
         }
 
         // ── Request line ─────────────────────────────────────────────────────
 
-        int methodEnd = find(buffer, base, headerEndIdx, (byte) ' ');
-        if (methodEnd == -1) throw new IOException("Invalid request line (method)");
+        int methodEnd = ByteScan.indexOf(buffer, base, headerEndIdx, (byte) ' ');
+        if (methodEnd == -1) throw new MalformedRequestException(400, "Invalid request line (method)");
+        if (methodEnd == base) throw new MalformedRequestException(400, "Missing HTTP method");
 
         HttpMethod method = HttpMethod.fromBytes(buffer, base, methodEnd - base);
-        if (method == null) throw new IOException("Unsupported HTTP method");
+        if (method == null) throw new MalformedRequestException(501, "Unsupported HTTP method");
 
         int pathStart = methodEnd + 1;
-        int pathEnd   = find(buffer, pathStart, headerEndIdx, (byte) ' ');
-        if (pathEnd == -1) throw new IOException("Invalid request line (path)");
+        int pathEnd   = ByteScan.indexOf(buffer, pathStart, headerEndIdx, (byte) ' ');
+        if (pathEnd == -1) throw new MalformedRequestException(400, "Invalid request line (path)");
 
-        int queryMark = find(buffer, pathStart, pathEnd, (byte) '?');
-        FastPathViews.RequestByteView pathView = new FastPathViews.RequestByteView(buffer, pathStart,
-                queryMark != -1 ? queryMark - pathStart : pathEnd - pathStart);
-        FastPathViews.RequestByteView queryView = queryMark != -1
-                ? new FastPathViews.RequestByteView(buffer, queryMark + 1, pathEnd - queryMark - 1)
-                : null;
+        int queryMark = ByteScan.indexOf(buffer, pathStart, pathEnd, (byte) '?');
+        pathView.reset(buffer, pathStart, queryMark != -1 ? queryMark - pathStart : pathEnd - pathStart);
+        if (queryMark != -1) queryView.reset(buffer, queryMark + 1, pathEnd - queryMark - 1);
 
         int protocolStart = pathEnd + 1;
-        int protocolEnd   = find(buffer, protocolStart, headerEndIdx, (byte) '\r');
-        if (protocolEnd == -1) throw new IOException("Invalid request line (protocol)");
+        int protocolEnd   = ByteScan.indexOf(buffer, protocolStart, headerEndIdx, (byte) '\r');
+        if (protocolEnd == -1) throw new MalformedRequestException(400, "Invalid request line (protocol)");
 
-        FastPathViews.RequestByteView protocolView =
-                new FastPathViews.RequestByteView(buffer, protocolStart, protocolEnd - protocolStart);
+        // from the overall header-block size, so an oversized request line gets its own,
+        // specific rejection rather than being folded into the generic "headers too large" case.
+        if (protocolEnd - base > Http1Limits.MAX_REQUEST_LINE_LENGTH) {
+            throw new MalformedRequestException(431, "Request line exceeds " + Http1Limits.MAX_REQUEST_LINE_LENGTH + " bytes");
+        }
+
+        protocolView.reset(buffer, protocolStart, protocolEnd - protocolStart);
 
         // ── Headers ──────────────────────────────────────────────────────────
 
-        int sectionStart = find(buffer, protocolEnd, headerEndIdx, (byte) '\n') + 1;
+        int sectionStart = ByteScan.indexOf(buffer, protocolEnd, headerEndIdx, (byte) '\n') + 1;
         int current      = sectionStart;
-        long contentLength = 0;
-        boolean isChunked  = false;
+        long contentLength          = -1;
+        boolean contentLengthSeen   = false;
+        boolean transferEncodingSeen    = false;
+        boolean transferEncodingChunked = false;
+        int headerCount = 0;
+        headerMap.beginParsed(buffer, sectionStart, headerEndIdx);
 
         while (current < headerEndIdx) {
-            int lineEnd = find(buffer, current, headerEndIdx + 1, (byte) '\r');
+            // deprecates line folding and treating a folded continuation as part of the
+            // previous header's value is a known request-smuggling vector.
+            byte first = buffer[current];
+            if (first == ' ' || first == '\t') {
+                throw new MalformedRequestException(400, "Obsolete line folding is not supported");
+            }
+
+            int lineEnd = ByteScan.indexOf(buffer, current, headerEndIdx + 1, (byte) '\r');
             if (lineEnd == -1 || lineEnd == current) break;
 
-            int colon = find(buffer, current, lineEnd, (byte) ':');
-            if (colon != -1) {
-                int valueStart = colon + 1;
-                while (valueStart < lineEnd && buffer[valueStart] == ' ') valueStart++;
+            // advancing past two bytes — a bare '\r' not followed by '\n' desynchronizes the
+            // parse and is a known bare-CR smuggling surface. Safe to read lineEnd+1: lineEnd
+            // is at most headerEndIdx, and findEndOfHeader already guaranteed 4 readable bytes
+            // (\r\n\r\n) starting at headerEndIdx.
+            if (buffer[lineEnd + 1] != '\n') {
+                throw new MalformedRequestException(400, "Malformed line terminator (bare CR)");
+            }
 
-                if (equalsIgnoreCase(buffer, current, colon, "content-length")) {
-                    contentLength = parseLong(buffer, valueStart, lineEnd);
-                } else if (equalsIgnoreCase(buffer, current, colon, "transfer-encoding")) {
-                    isChunked = equalsIgnoreCase(buffer, valueStart, lineEnd, "chunked");
+            if (++headerCount > Http1Limits.MAX_HEADER_COUNT) {
+                throw new MalformedRequestException(431, "Too many headers");
+            }
+
+            int colon = ByteScan.indexOf(buffer, current, lineEnd, (byte) ':');
+            if (colon == -1) {
+                throw new MalformedRequestException(400, "Header line missing ':'");
+            }
+            if (colon - current > Http1Limits.MAX_HEADER_NAME_LENGTH) {
+                throw new MalformedRequestException(431, "Header name exceeds " + Http1Limits.MAX_HEADER_NAME_LENGTH + " bytes");
+            }
+            for (int i = current; i < colon; i++) {
+                if (!ByteScan.isTChar(buffer[i])) {
+                    throw new MalformedRequestException(400, "Invalid header name character");
                 }
             }
+
+            int valueStart = colon + 1;
+            while (valueStart < lineEnd && buffer[valueStart] == ' ') valueStart++;
+            if (lineEnd - valueStart > Http1Limits.MAX_HEADER_VALUE_LENGTH) {
+                throw new MalformedRequestException(431, "Header value exceeds " + Http1Limits.MAX_HEADER_VALUE_LENGTH + " bytes");
+            }
+            headerMap.addParsed(current, colon - current, valueStart, lineEnd - valueStart);
+
+            if (ByteScan.equalsIgnoreCaseAscii(buffer, current, colon, "content-length")) {
+                // parseLong, which silently accepted "5abc" as 5 and "-1" as 1.
+                long parsed = parseContentLengthStrict(buffer, valueStart, lineEnd);
+                // Multiple Content-Length lines with differing values is itself a smuggling
+                // permits a recipient to treat that as one value).
+                if (contentLengthSeen && parsed != contentLength) {
+                    throw new MalformedRequestException(400, "Conflicting Content-Length values");
+                }
+                contentLength = parsed;
+                contentLengthSeen = true;
+            } else if (ByteScan.equalsIgnoreCaseAscii(buffer, current, colon, "transfer-encoding")) {
+                transferEncodingSeen = true;
+                // "chunked", so "gzip, chunked" — valid per RFC 9112 §6.1, where chunked need
+                // only be the FINAL coding — was silently treated as not chunked at all,
+                // corrupting the message boundary. Fixed by inspecting only the last token.
+                transferEncodingChunked = isFinalCodingChunked(buffer, valueStart, lineEnd);
+            }
             current = lineEnd + 2;
         }
 
-        headerMap.reset(buffer, sectionStart, headerEndIdx);
+        // MUST be treated as an error by an origin server — this is the canonical CL.TE/TE.CL
+        // smuggling vector. Checked once both headers are known, regardless of the order they
+        // appeared in, so ordering games cannot bypass it.
+        if (contentLengthSeen && transferEncodingSeen) {
+            throw new MalformedRequestException(400, "Content-Length and Transfer-Encoding both present");
+        }
+        boolean isChunked;
+        if (transferEncodingSeen) {
+            if (!transferEncodingChunked) {
+                throw new MalformedRequestException(501, "Unsupported Transfer-Encoding");
+            }
+            isChunked = true;
+        } else {
+            isChunked = false;
+            if (!contentLengthSeen) contentLength = 0;
+        }
 
         // ── Body / pipelining accounting ─────────────────────────────────────
 
@@ -181,50 +290,69 @@ public class RequestParser {
             preBufLen = (int) contentLength;
         }
 
-        RequestLine requestLine = new RequestLine(method, pathView, queryView, protocolView, headerMap);
+        requestLine.reset(method, pathView, queryMark != -1 ? queryView : null, protocolView, headerMap);
 
+        // Javadoc) -- reset() repositions it for the fixed-length/empty case (contentLength == 0
+        // is handled by the same call: preBufLen is already forced to 0 for it above) or the
+        // chunked case, never reallocated.
         if (isChunked) {
-            return Request.forParsed(requestLine,
-                    new ChunkedInputStream(in, buffer, bodyStart, preBufLen),
-                    -1L, null, 0, 0, remoteAddress, sslSocket);
+            requestBody.reset(
+                    new ChunkedInputStream(in, buffer, bodyStart, preBufLen, trailerMap),
+                    -1L, null, 0, 0);
+        } else {
+            requestBody.reset(in, contentLength, buffer, bodyStart, preBufLen);
         }
-        return Request.forParsed(requestLine, in, contentLength, buffer, bodyStart, preBufLen, remoteAddress, sslSocket);
+        Request parsed = Request.forParsed(request, requestLine, requestBody, remoteAddress, sslSocket);
+        parsed.setTrailers(trailerMap);
+        return parsed;
     }
 
-    // ── Buffer scanning utilities (hot path — keep branch-free where possible) ──
-
-    private static int findEndOfHeader(byte[] buf, int from, int len) {
-        for (int i = from; i <= len - 4; i++) {
-            if (buf[i] == '\r' && buf[i+1] == '\n' && buf[i+2] == '\r' && buf[i+3] == '\n')
-                return i;
-        }
-        return -1;
-    }
-
-    private static int find(byte[] buf, int start, int end, byte target) {
-        for (int i = start; i < end; i++) {
-            if (buf[i] == target) return i;
-        }
-        return -1;
-    }
-
-    private static boolean equalsIgnoreCase(byte[] buf, int start, int end, String target) {
+    /**
+     * value, any non-digit byte (including a leading {@code +}/{@code -}, which are not
+     * digits), more than 19 digits (the longest possible {@code Long.MAX_VALUE}), arithmetic
+     * overflow past {@code Long.MAX_VALUE}, and a value above
+     * {@link Http1Limits#MAX_CONTENT_LENGTH}. The pre-existing {@code parseLong} silently
+     * skipped any non-digit character instead of rejecting it — {@code "5abc"} parsed as
+     * {@code 5} and {@code "-1"} parsed as {@code 1}.
+     */
+    private static long parseContentLengthStrict(byte[] buf, int start, int end) throws MalformedRequestException {
         int len = end - start;
-        if (len != target.length()) return false;
-        for (int i = 0; i < len; i++) {
-            byte b = buf[start + i];
-            if (b >= 'A' && b <= 'Z') b += 32;
-            if (b != (byte) target.charAt(i)) return false;
-        }
-        return true;
-    }
-
-    private static long parseLong(byte[] buf, int start, int end) {
+        if (len == 0) throw new MalformedRequestException(400, "Empty Content-Length value");
+        if (len > 19) throw new MalformedRequestException(400, "Content-Length value too long");
         long value = 0;
         for (int i = start; i < end; i++) {
             byte c = buf[i];
-            if (c >= '0' && c <= '9') value = value * 10 + (c - '0');
+            if (c < '0' || c > '9') {
+                throw new MalformedRequestException(400, "Malformed Content-Length value");
+            }
+            int digit = c - '0';
+            if (value > (Long.MAX_VALUE - digit) / 10) {
+                throw new MalformedRequestException(400, "Content-Length overflow");
+            }
+            value = value * 10 + digit;
+        }
+        if (value > Http1Limits.MAX_CONTENT_LENGTH) {
+            throw new MalformedRequestException(413, "Content-Length exceeds configured maximum");
         }
         return value;
     }
+
+    /**
+     * RFC 9112 §6.1: when {@code Transfer-Encoding} lists multiple codings
+     * ({@code "gzip, chunked"}), {@code chunked} MUST be the final one for the message to be
+     * self-delimiting. Returns whether the last comma-separated token in {@code [start, end)}
+     * is exactly {@code "chunked"} (case-insensitive), ignoring surrounding whitespace around
+     * misclassified any multi-coding value as non-chunked.
+     */
+    private static boolean isFinalCodingChunked(byte[] buf, int start, int end) {
+        int e = end;
+        while (e > start && (buf[e - 1] == ' ' || buf[e - 1] == '\t')) e--;
+        int lastComma = start - 1;
+        for (int i = start; i < e; i++) {
+            if (buf[i] == ',') lastComma = i;
+        }
+        int tokenStart = lastComma + 1;
+        while (tokenStart < e && (buf[tokenStart] == ' ' || buf[tokenStart] == '\t')) tokenStart++;
+        return ByteScan.equalsIgnoreCaseAscii(buf, tokenStart, e, "chunked");
+    }
 }

flash/src/main/java/dev/relism/flash/ServerHandle.java

@@ -4,6 +4,7 @@ import dev.relism.flash.extension.FlashApp;
 import dev.relism.flash.extension.FlashConfiguration;
 import dev.relism.flash.routing.AbstractRouter;
 import dev.relism.flash.routing.AbstractWsRouter;
+import dev.relism.flash.transport.TransportFactory;
 
 import java.io.IOException;
 import java.util.concurrent.CompletableFuture;
@@ -11,7 +12,7 @@ import java.util.concurrent.CompletableFuture;
 /**
  * Public handle to the underlying HTTP transport. Returned by {@link #create}
  * so that {@link FlashApp} can start and stop the server
- * without holding a direct reference to the package-private {@link HttpServer}.
+ * without holding a direct reference to the transport's internal composition
  */
 public interface ServerHandle {
 
@@ -30,6 +31,6 @@ public interface ServerHandle {
     static ServerHandle create(FlashConfiguration config,
                                AbstractRouter httpRouter,
                                AbstractWsRouter wsRouter) throws IOException {
-        return new HttpServer(config, httpRouter, wsRouter);
+        return TransportFactory.create(config, httpRouter, wsRouter);
     }
 }

flash/src/main/java/dev/relism/flash/api/multipart/Multipart.java

@@ -1,7 +1,9 @@
 package dev.relism.flash.api.multipart;
 
+import dev.relism.flash.http.Http1Limits;
 import dev.relism.flash.models.Request;
 
+import java.io.ByteArrayOutputStream;
 import java.io.IOException;
 import java.io.InputStream;
 import java.nio.charset.StandardCharsets;
@@ -41,6 +43,7 @@ import java.util.*;
  * <p><b>Thread safety:</b> not thread-safe; one instance per request.
  */
 public final class Multipart {
+    private int partCount;
 
     private static final int BUF_CAP = 8192;
 
@@ -156,6 +159,11 @@ public final class Multipart {
         Map<String, String> headers = readPartHeaders();
         if (headers == null) { done = true; return null; }
 
+        // unbounded growth of `scanned` and unbounded cumulative header-parsing work.
+        if (++partCount > Http1Limits.MAX_MULTIPART_PARTS) {
+            throw new IOException("multipart body exceeds max part count (" + Http1Limits.MAX_MULTIPART_PARTS + ")");
+        }
+
         String disp     = headers.get("content-disposition");
         String name     = extractParam(disp, "name");
         String filename = extractParam(disp, "filename");
@@ -168,8 +176,9 @@ public final class Multipart {
             // File part — expose streaming body; not cached (stream is consumed once)
             p = Part.streaming(name, filename, ct, active);
         } else {
-            // Text part, or full-scan path: buffer body now
-            byte[] body = active.readAllBytes();
+            // InputStream.readAllBytes() — an unbounded field/file body would otherwise let a
+            // hostile peer force an arbitrarily large single heap allocation.
+            byte[] body = readBoundedBody(active);
             active = null;
             p = Part.buffered(name, filename, ct, body);
             scanned.add(p);
@@ -177,6 +186,27 @@ public final class Multipart {
         return p;
     }
 
+    /**
+     * Reads {@code in} to EOF into a {@code byte[]}, bounded by
+     * {@link Http1Limits#MAX_MULTIPART_BUFFERED_PART_SIZE} — see that constant's Javadoc for why
+     * this bound is necessary even though the overall request body already has one.
+     */
+    private static byte[] readBoundedBody(InputStream in) throws IOException {
+        ByteArrayOutputStream out = new ByteArrayOutputStream(BUF_CAP);
+        byte[] chunk = new byte[BUF_CAP];
+        long total = 0;
+        int n;
+        while ((n = in.read(chunk)) > 0) {
+            total += n;
+            if (total > Http1Limits.MAX_MULTIPART_BUFFERED_PART_SIZE) {
+                throw new IOException("multipart part body exceeds max buffered size ("
+                        + Http1Limits.MAX_MULTIPART_BUFFERED_PART_SIZE + " bytes)");
+            }
+            out.write(chunk, 0, n);
+        }
+        return out.toByteArray();
+    }
+
     // -------------------------------------------------------------------------
     // PartBodyStream — inner class sharing the window buffer
     // -------------------------------------------------------------------------
@@ -266,9 +296,15 @@ public final class Multipart {
 
     private Map<String, String> readPartHeaders() throws IOException {
         Map<String, String> map = new HashMap<>();
+        int count = 0;
         while (true) {
             String line = readLine();
             if (line == null || line.isEmpty()) break;
+            // header lines before the blank line that ends a part's header block.
+            if (++count > Http1Limits.MAX_MULTIPART_PART_HEADER_COUNT) {
+                throw new IOException("multipart part exceeds max header count ("
+                        + Http1Limits.MAX_MULTIPART_PART_HEADER_COUNT + ")");
+            }
             int colon = line.indexOf(':');
             if (colon > 0)
                 map.put(line.substring(0, colon).trim().toLowerCase(Locale.ROOT),
@@ -289,6 +325,7 @@ public final class Multipart {
                     sb.append(new String(win, wPos, i - wPos, StandardCharsets.UTF_8));
                     int consumed = i - wPos + 2;
                     wPos += consumed; wLen -= consumed;
+                    checkHeaderLineLength(sb.length());
                     return sb.toString();
                 }
             }
@@ -298,14 +335,26 @@ public final class Multipart {
                 sb.append(new String(win, wPos, append, StandardCharsets.UTF_8));
                 wPos += append; wLen -= append;
             }
+            // growing for as long as it keeps streaming bytes — the multipart-header analogue of
+            // RequestParser's Http1Limits.MAX_HEADER_VALUE_LENGTH check, which does not apply
+            // here since these header lines live inside the body, not the top-level HTTP headers.
+            checkHeaderLineLength(sb.length());
             if (srcEof && wLen > 0) {
                 sb.append(new String(win, wPos, wLen, StandardCharsets.UTF_8));
                 wPos += wLen; wLen = 0;
+                checkHeaderLineLength(sb.length());
                 return sb.toString();
             }
         }
     }
 
+    private static void checkHeaderLineLength(int length) throws IOException {
+        if (length > Http1Limits.MAX_MULTIPART_HEADER_LINE_LENGTH) {
+            throw new IOException("multipart header line exceeds "
+                    + Http1Limits.MAX_MULTIPART_HEADER_LINE_LENGTH + " bytes");
+        }
+    }
+
     // -------------------------------------------------------------------------
     // Utilities
     // -------------------------------------------------------------------------

flash/src/main/java/dev/relism/flash/bytes/ArrayBackedByteView.java

@@ -0,0 +1,36 @@
+package dev.relism.flash.bytes;
+
+import dev.relism.fpr.core.ByteView;
+
+/**
+ * Capability interface for a {@link ByteView} that is a contiguous slice of a single backing
+ * {@code byte[]} — as opposed to a {@link SegmentedByteView}, which spans several arrays and
+ * cannot expose a single {@code (array, offset)} pair.
+ *
+ * <p>Every array-backed view in this codebase implements this: {@code RequestByteView},
+ * {@code SocketByteView}, {@code StringByteView} (all in
+ * {@code dev.relism.flash.routing.routers.fastpathrouter.FastPathViews}), and {@link PooledSlice}.
+ * {@code MethodPathByteView} deliberately does not — it is a composite of a {@code byte[]}
+ * (method) and another {@link ByteView} (path), so it has no single backing array.
+ *
+ * <h3>What this enables</h3>
+ * Anywhere code holds a plain {@link ByteView} and wants the fast path when the concrete
+ * instance happens to be array-backed, an {@code instanceof ArrayBackedByteView} check unlocks:
+ * <ul>
+ *   <li>Single-allocation {@code String} construction —
+ *       {@code new String(view.array(), view.offset(), view.length(), UTF_8)} instead of a
+ *       byte-at-a-time copy into a scratch {@code byte[]} followed by a second allocation for
+ *   <li>A single {@code System.arraycopy} instead of a manual loop wherever a view's bytes need
+ *       to be copied.</li>
+ * </ul>
+ * Code that only has a bare {@link ByteView} (e.g. because it received one across the
+ * {@link SegmentedByteView} boundary) keeps the
+ * byte-at-a-time fallback — this interface is an opportunistic fast path, never a requirement.
+ */
+public interface ArrayBackedByteView extends ByteView {
+    /** The backing array. Bytes {@code [offset(), offset() + length())} belong to this view. */
+    byte[] array();
+
+    /** Offset of this view's first byte within {@link #array()}. */
+    int offset();
+}

flash/src/main/java/dev/relism/flash/bytes/ByteScan.java

@@ -0,0 +1,324 @@
+package dev.relism.flash.bytes;
+
+import dev.relism.fpr.core.ByteView;
+
+import java.lang.invoke.MethodHandles;
+import java.lang.invoke.VarHandle;
+import java.nio.ByteOrder;
+
+/**
+ * The single home for protocol-neutral byte scanning: single-byte search, the four-byte
+ * {@code \r\n\r\n} header-terminator search (SWAR-accelerated), case-insensitive comparison,
+ * comma-separated token-list scanning ({@code Connection: a, b, c}), RFC 9110 {@code tchar}
+ * validation, and the case-insensitive header-name hash {@link dev.relism.flash.models.Http1HeaderMap}'s
+ *
+ * <p>Every method here is {@code static} and allocates nothing. Every SWAR method has a plain
+ * scalar counterpart ({@code *Scalar}) that exists for two reasons: it is what the tests use as
+ * the correctness oracle (property-tested against the SWAR version on randomized inputs — see
+ * {@code ByteScanTest}/{@code ByteScanFuzzTest}), and it is the documented fallback if a future
+ * measurement ever shows the SWAR path is not worth its complexity on some path (none has been
+ *
+ * <h3>The SWAR technique used throughout</h3>
+ * Both {@link #indexOf} and {@link #indexOfCrLfCrLf} use the classic "does this word contain
+ * byte {@code b}" bit trick (Bit Twiddling Hacks, "Determine if a word has a byte equal to n"):
+ * XOR the 8-byte word against {@code b} broadcast into every lane (turning matching lanes to
+ * {@code 0x00}), then test for any zero lane with
+ * {@code (v - 0x0101010101010101L) & ~v & 0x8080808080808080L} — non-zero exactly when some lane
+ * was {@code 0x00} before the subtraction, i.e. some original lane equalled {@code b}. This finds
+ * *that a* matching lane exists in one word-sized read plus a handful of ALU ops, touching every
+ * byte only once per 8-byte stride in the common (no-match-yet) case, instead of once per byte.
+ *
+ * <p>Reading the word uses {@link MethodHandles#byteArrayViewVarHandle} with
+ * {@link ByteOrder#nativeOrder()} — deliberately native rather than a fixed order (contrast
+ * {@code fpr-core}'s {@code ByteCompare}, which fixes {@code LITTLE_ENDIAN} because it compares
+ * two independently-read words for bit-exact equality and so needs a byte order both reads
+ * agree on; nothing here compares across two separately-decoded words, so the fastest order for
+ * the host CPU is free to use). Byte-equality detection itself (finding that a matching lane
+ * exists in the mask) does not depend on which order was used to assemble the word — XOR and the
+ * haszero test are lane-wise operations, indifferent to how lanes map to memory offsets.
+ * <b>Position extraction does depend on it</b>: converting "which bit of the 64-bit mask is set"
+ * back into "which array index did that byte come from" requires knowing whether array byte 0
+ * became the long's least-significant byte (little-endian) or most-significant byte
+ * (big-endian) — {@link #laneIndexOf} branches on {@link #NATIVE_IS_LITTLE} once, at class-init
+ * time, precisely to get this right on either host.
+ */
+public final class ByteScan {
+    private ByteScan() {}
+
+    private static final ByteOrder NATIVE_ORDER = ByteOrder.nativeOrder();
+    private static final boolean NATIVE_IS_LITTLE = NATIVE_ORDER == ByteOrder.LITTLE_ENDIAN;
+    private static final VarHandle LONG_VIEW =
+            MethodHandles.byteArrayViewVarHandle(long[].class, NATIVE_ORDER);
+
+    private static final long LANE_LSB = 0x0101010101010101L;
+    private static final long LANE_MSB = 0x8080808080808080L;
+
+    // ── tchar (RFC 9110 §5.6.2) ──────────────────────────────────────────────
+
+    /**
+     * RFC 9110 §5.6.2 {@code tchar} set, table-driven so validation is a single array read per
+     * only the ASCII range a valid header-name character can ever occupy is populated.
+     */
+    private static final boolean[] TCHAR = new boolean[128];
+
+    static {
+        for (byte b : "!#$%&'*+-.^_`|~".getBytes(java.nio.charset.StandardCharsets.US_ASCII)) {
+            TCHAR[b] = true;
+        }
+        for (char c = '0'; c <= '9'; c++) TCHAR[c] = true;
+        for (char c = 'A'; c <= 'Z'; c++) TCHAR[c] = true;
+        for (char c = 'a'; c <= 'z'; c++) TCHAR[c] = true;
+    }
+
+    /** Whether {@code b} is a valid RFC 9110 §5.6.2 {@code tchar} (a legal header-name byte). */
+    public static boolean isTChar(byte b) {
+        return b >= 0 && b < 128 && TCHAR[b];
+    }
+
+    // ── Single-byte search ───────────────────────────────────────────────────
+
+    /**
+     * Index of the first occurrence of {@code target} in {@code buf[from, to)}, or {@code -1}.
+     * SWAR-accelerated: touches 8 bytes per word while no match has been found, falling back to
+     * a byte-at-a-time tail once fewer than 8 bytes remain.
+     */
+    public static int indexOf(byte[] buf, int from, int to, byte target) {
+        long broadcast = (target & 0xFFL) * LANE_LSB;
+        int i = from;
+        while (i + 8 <= to) {
+            long word = (long) LONG_VIEW.get(buf, i);
+            long masked = hasZeroLane(word ^ broadcast);
+            if (masked != 0) {
+                return i + laneIndexOf(masked);
+            }
+            i += 8;
+        }
+        for (; i < to; i++) {
+            if (buf[i] == target) return i;
+        }
+        return -1;
+    }
+
+    /** Plain byte-at-a-time reference implementation of {@link #indexOf} — the test oracle. */
+    static int indexOfScalar(byte[] buf, int from, int to, byte target) {
+        for (int i = from; i < to; i++) {
+            if (buf[i] == target) return i;
+        }
+        return -1;
+    }
+
+    // ── \r\n\r\n header terminator search ────────────────────────────────────
+
+    private static final byte CR = '\r', LF = '\n';
+
+    /**
+     * Index of the first {@code "\r\n\r\n"} in {@code buf[from, to)}, or {@code -1}. SWAR
+     * pre-filter (find a candidate {@code CR} byte 8 at a time) plus a cheap scalar 3-byte
+     * verify at each candidate — see the class Javadoc for the technique and
+     */
+    public static int indexOfCrLfCrLf(byte[] buf, int from, int to) {
+        int limit = to - 4; // last index at which a 4-byte match can start
+        int i = from;
+        while (i + 8 <= to) {
+            long word = (long) LONG_VIEW.get(buf, i);
+            long masked = hasZeroLane(word ^ CR_BROADCAST);
+            if (masked == 0) {
+                i += 8;
+                continue;
+            }
+            int crPos = i + laneIndexOf(masked);
+            if (crPos > limit) {
+                // Nearest CR candidate in this word can't fit a full match before `to`; no CR
+                // exists before it in [i, crPos) (laneIndexOf always finds the lowest-address
+                // match first), so nothing in [i, crPos) can match either — the scalar tail
+                // below, bounded by `limit`, correctly finds nothing without re-deriving that.
+                break;
+            }
+            if (buf[crPos + 1] == LF && buf[crPos + 2] == CR && buf[crPos + 3] == LF) {
+                return crPos;
+            }
+            i = crPos + 1;
+        }
+        for (; i <= limit; i++) {
+            if (buf[i] == CR && buf[i + 1] == LF && buf[i + 2] == CR && buf[i + 3] == LF) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    private static final long CR_BROADCAST = (CR & 0xFFL) * LANE_LSB;
+
+    /** Plain byte-at-a-time reference implementation of {@link #indexOfCrLfCrLf} — the test oracle. */
+    static int indexOfCrLfCrLfScalar(byte[] buf, int from, int to) {
+        for (int i = from; i <= to - 4; i++) {
+            if (buf[i] == CR && buf[i + 1] == LF && buf[i + 2] == CR && buf[i + 3] == LF) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    /** "Determine if a word has a byte equal to n" (Bit Twiddling Hacks), applied to {@code xored}. */
+    private static long hasZeroLane(long xored) {
+        return (xored - LANE_LSB) & ~xored & LANE_MSB;
+    }
+
+    /** Converts a {@link #hasZeroLane} result into the array-index offset of its lowest matching lane. */
+    private static int laneIndexOf(long masked) {
+        return NATIVE_IS_LITTLE
+                ? Long.numberOfTrailingZeros(masked) >>> 3
+                : 7 - (Long.numberOfLeadingZeros(masked) >>> 3);
+    }
+
+    // ── Case-insensitive comparison ──────────────────────────────────────────
+
+    private static byte foldAsciiUpper(byte b) {
+        return (b >= 'A' && b <= 'Z') ? (byte) (b + 32) : b;
+    }
+
+    /** Case-insensitive (ASCII) equality of {@code buf[start, end)} against {@code target}. */
+    public static boolean equalsIgnoreCaseAscii(byte[] buf, int start, int end, String target) {
+        int len = end - start;
+        if (len != target.length()) return false;
+        for (int i = 0; i < len; i++) {
+            if (foldAsciiUpper(buf[start + i]) != foldAsciiUpper((byte) target.charAt(i))) return false;
+        }
+        return true;
+    }
+
+    /** Case-insensitive (ASCII) equality of two byte-array ranges. */
+    public static boolean equalsIgnoreCaseAscii(byte[] a, int aStart, int aLen, byte[] b, int bStart, int bLen) {
+        if (aLen != bLen) return false;
+        for (int i = 0; i < aLen; i++) {
+            if (foldAsciiUpper(a[aStart + i]) != foldAsciiUpper(b[bStart + i])) return false;
+        }
+        return true;
+    }
+
+    /** Case-insensitive (ASCII) equality of {@code view[start, end)} against {@code target}. */
+    public static boolean equalsIgnoreCase(ByteView view, int start, int end, String target) {
+        int len = end - start;
+        if (len != target.length()) return false;
+        for (int i = 0; i < len; i++) {
+            if (foldAsciiUpper(view.byteAt(start + i)) != foldAsciiUpper((byte) target.charAt(i))) return false;
+        }
+        return true;
+    }
+
+    // ── Comma-separated token lists (e.g. `Connection: keep-alive, Upgrade`) ────
+
+    /**
+     * Whether the comma-separated, OWS-tolerant token list {@code view} contains {@code token}
+     * (case-insensitive). The shared scanner behind both {@code Http1KeepAlive.isKeepAlive} and
+     * drift apart the way a whole-value {@code equals} check once did.
+     */
+    public static boolean tokenListContains(ByteView view, String token) {
+        int len = view.length(), i = 0;
+        while (i < len) {
+            while (i < len && view.byteAt(i) == ' ') i++;
+            int start = i;
+            while (i < len && view.byteAt(i) != ',') i++;
+            if (tokenEqualsIgnoreCase(view, start, i, token)) return true;
+            i++;
+        }
+        return false;
+    }
+
+    /** Case-insensitive compare of {@code view[start, end)}, trimming trailing spaces, against {@code token}. */
+    public static boolean tokenEqualsIgnoreCase(ByteView view, int start, int end, String token) {
+        int wlen = end - start;
+        while (wlen > 0 && view.byteAt(start + wlen - 1) == ' ') wlen--;
+        return equalsIgnoreCase(view, start, start + wlen, token);
+    }
+
+
+    /**
+     * Case-insensitive (ASCII fold) 32-bit FNV-1a hash of {@code buf[start, start + len)}. Used
+     * by {@link dev.relism.flash.models.Http1HeaderMap}'s per-request index to compare a cheap hash
+     * before falling back to a full case-insensitive {@code memcmp}-equivalent
+     * ({@link #equalsIgnoreCaseAscii}) — two header names that differ anywhere hash differently
+     * with overwhelming probability, so the common "not the header I'm looking for" case resolves
+     * in one hash compare instead of a byte-by-byte scan.
+     */
+    public static int hashNameIgnoreCaseAscii(byte[] buf, int start, int len) {
+        int hash = 0x811C9DC5; // FNV-1a 32-bit offset basis
+        for (int i = 0; i < len; i++) {
+            hash ^= (foldAsciiUpper(buf[start + i]) & 0xFF);
+            hash *= 0x01000193; // FNV-1a 32-bit prime
+        }
+        return hash;
+    }
+
+    /**
+     * Same hash as {@link #hashNameIgnoreCaseAscii(byte[], int, int)}, computed directly from a
+     * lookup-key {@code String} (e.g. {@code "Content-Type"}) instead of already-scanned bytes —
+     * the two must agree bit-for-bit on equivalent ASCII content for
+     * {@link dev.relism.flash.models.Http1HeaderMap}'s index (hash the request-declared bytes once at
+     * {@code reset()}; hash the caller's lookup key once per {@code first()}/{@code all()} call;
+     * compare the two cheap hashes before ever touching a full case-insensitive comparison).
+     */
+    public static int hashNameIgnoreCaseAscii(String name) {
+        int hash = 0x811C9DC5;
+        int len = name.length();
+        for (int i = 0; i < len; i++) {
+            hash ^= (foldAsciiUpper((byte) name.charAt(i)) & 0xFF);
+            hash *= 0x01000193;
+        }
+        return hash;
+    }
+
+    // ── Decimal / hex parsing ────────────────────────────────────────────────
+
+    /** Sentinel returned by {@link #parseDecimalStrict} on any malformed or out-of-range input. */
+    public static final long PARSE_INVALID = -1L;
+
+    /**
+     * Strict, overflow-safe unsigned decimal parse of {@code buf[start, end)}: rejects an empty
+     * range, any non-{@code '0'..'9'} byte, more than 19 digits, and arithmetic overflow past
+     * {@link Long#MAX_VALUE}. Returns {@link #PARSE_INVALID} rather than throwing — the same
+     * shape {@code RequestParser}'s own {@code Content-Length} parser already hand-rolls (kept
+     * separate there since it also needs to throw a specific, differently-worded
+     * {@code MalformedRequestException} per failure mode); this is the general-purpose version
+     * for callers (HPACK integer decoding, frame-length fields) that just need a valid/invalid
+     * signal.
+     */
+    public static long parseDecimalStrict(byte[] buf, int start, int end) {
+        int len = end - start;
+        if (len == 0 || len > 19) return PARSE_INVALID;
+        long value = 0;
+        for (int i = start; i < end; i++) {
+            byte c = buf[i];
+            if (c < '0' || c > '9') return PARSE_INVALID;
+            int digit = c - '0';
+            if (value > (Long.MAX_VALUE - digit) / 10) return PARSE_INVALID;
+            value = value * 10 + digit;
+        }
+        return value;
+    }
+
+    /**
+     * Parses up to {@code maxDigits} hex digits (ASCII, either case) from {@code buf[start, end)}
+     * as an unsigned value. Returns {@link #PARSE_INVALID} if the range is empty, contains a
+     * non-hex-digit byte, or would need more than {@code maxDigits} digits to represent (the
+     * caller's bound against, e.g., a chunk-size line with an implausible number of digits).
+     */
+    public static long parseHexStrict(byte[] buf, int start, int end, int maxDigits) {
+        int len = end - start;
+        if (len == 0 || len > maxDigits) return PARSE_INVALID;
+        long value = 0;
+        for (int i = start; i < end; i++) {
+            int digit = hexDigit(buf[i]);
+            if (digit < 0) return PARSE_INVALID;
+            value = (value << 4) | digit;
+        }
+        return value;
+    }
+
+    private static int hexDigit(byte b) {
+        if (b >= '0' && b <= '9') return b - '0';
+        if (b >= 'a' && b <= 'f') return b - 'a' + 10;
+        if (b >= 'A' && b <= 'F') return b - 'A' + 10;
+        return -1;
+    }
+}

flash/src/main/java/dev/relism/flash/bytes/ByteWriter.java

@@ -0,0 +1,162 @@
+package dev.relism.flash.bytes;
+
+import java.nio.charset.StandardCharsets;
+
+/**
+ * Index-based writer into a growable {@code byte[]} scratch buffer. Every {@code write*} method
+ * bounds-checks and grows the backing array only when the write would not otherwise fit —
+ * on an already-warm buffer (the steady-state case: the buffer has already grown to the
+ * connection's high-water mark), no method here allocates.
+ *
+ * Callers build a complete message in a {@code ByteWriter}-backed scratch buffer and then issue
+ * one bulk {@code write(buffer, 0, length())}. The same writer is shared by HTTP/1.1 and HTTP/2.
+ *
+ * <h3>Lifetime and thread-safety contract</h3>
+ * Not thread-safe — exactly one writer at a time, matching every other per-connection scratch
+ * object in this codebase ({@code ConnectionScratch}, {@code Http1HeaderMap}). {@link #reset()}
+ * repositions this writer to the start of its backing array for the next message; the backing
+ * array itself is never shrunk back down, only grown — the same amortized-to-zero-allocation
+ * growth policy {@code RequestParser}'s read buffer already uses.
+ */
+public final class ByteWriter {
+    private byte[] buf;
+    private final byte[] digits = new byte[20];
+    private int len;
+
+    public ByteWriter(int initialCapacity) {
+        this.buf = new byte[Math.max(initialCapacity, 16)];
+    }
+
+    /** Repositions this writer to the start of its buffer, ready for the next message. */
+    public void reset() {
+        len = 0;
+    }
+
+    /** The backing buffer. Valid content is {@code [0, length())} — never assume {@code buf.length == length()}. */
+    public byte[] array() {
+        return buf;
+    }
+
+    /** How many bytes have been written since the last {@link #reset()}. */
+    public int length() {
+        return len;
+    }
+
+    private void ensure(int additional) {
+        int needed = len + additional;
+        if (needed <= buf.length) return;
+        int grown = buf.length * 2;
+        while (grown < needed) grown *= 2;
+        byte[] next = new byte[grown];
+        System.arraycopy(buf, 0, next, 0, len);
+        buf = next;
+    }
+
+    public void writeByte(byte b) {
+        ensure(1);
+        buf[len++] = b;
+    }
+
+    public void writeBytes(byte[] src) {
+        writeBytes(src, 0, src.length);
+    }
+
+    public void writeBytes(byte[] src, int off, int srcLen) {
+        ensure(srcLen);
+        System.arraycopy(src, off, buf, len, srcLen);
+        len += srcLen;
+    }
+
+    /**
+     * Writes {@code value}'s ASCII decimal digits (no sign — callers write {@code '-'} via
+     * {@link #writeByte} first if needed). {@code value} must be non-negative.
+     */
+    public void writeDecimal(long value) {
+        if (value < 0) throw new IllegalArgumentException("writeDecimal requires a non-negative value: " + value);
+        if (value == 0) {
+            writeByte((byte) '0');
+            return;
+        }
+        // Digits emerge least-significant-first. The reusable field holds every possible long
+        // representation, so decimal rendering does not allocate on a warm writer.
+        int n = 0;
+        long v = value;
+        while (v > 0) {
+            digits[n++] = (byte) ('0' + (v % 10));
+            v /= 10;
+        }
+        ensure(n);
+        for (int i = n - 1; i >= 0; i--) buf[len++] = digits[i];
+    }
+
+    private static final byte[] HEX_DIGITS = "0123456789abcdef".getBytes(StandardCharsets.US_ASCII);
+
+    /** Writes {@code value}'s lowercase hex digits, no leading zeros (except for {@code value == 0}, which writes {@code "0"}). */
+    public void writeHex(int value) {
+        if (value == 0) {
+            writeByte((byte) '0');
+            return;
+        }
+        int n = 0;
+        int v = value;
+        while (v != 0) {
+            digits[n++] = HEX_DIGITS[v & 0xF];
+            v >>>= 4;
+        }
+        ensure(n);
+        for (int i = n - 1; i >= 0; i--) buf[len++] = digits[i];
+    }
+
+    /** Writes {@code s}'s ASCII bytes, lower-cased. {@code s} must be ASCII-only. */
+    public void writeAsciiLower(String s) {
+        int n = s.length();
+        ensure(n);
+        for (int i = 0; i < n; i++) {
+            char c = s.charAt(i);
+            if (c >= 'A' && c <= 'Z') c += 32;
+            buf[len++] = (byte) c;
+        }
+    }
+
+    /**
+     * Writes {@code s}'s ASCII bytes, case preserved. {@code s} must be ASCII-only. Unlike
+     * {@code new String(...).getBytes(UTF_8)}, writes each character directly into this buffer
+     * and avoids an intermediate {@code byte[]}.
+     */
+    public void writeAscii(String s) {
+        int n = s.length();
+        ensure(n);
+        for (int i = 0; i < n; i++) {
+            buf[len++] = (byte) s.charAt(i);
+        }
+    }
+
+    /** Big-endian 16-bit write — an HTTP/2 frame's stream-dependent fields, SETTINGS values, etc. */
+    public void writeUInt16(int value) {
+        ensure(2);
+        buf[len++] = (byte) (value >>> 8);
+        buf[len++] = (byte) value;
+    }
+
+    /** Big-endian 24-bit write — an HTTP/2 frame header's length field. */
+    public void writeUInt24(int value) {
+        ensure(3);
+        buf[len++] = (byte) (value >>> 16);
+        buf[len++] = (byte) (value >>> 8);
+        buf[len++] = (byte) value;
+    }
+
+    /** Big-endian 31-bit write (top bit always 0) — an HTTP/2 stream identifier. */
+    public void writeUInt31(int value) {
+        writeUInt32(value & 0x7FFFFFFF);
+    }
+
+    /** Big-endian 32-bit write — an HTTP/2 window-size increment, SETTINGS value, etc. */
+    public void writeUInt32(int value) {
+        ensure(4);
+        buf[len++] = (byte) (value >>> 24);
+        buf[len++] = (byte) (value >>> 16);
+        buf[len++] = (byte) (value >>> 8);
+        buf[len++] = (byte) value;
+    }
+}

flash/src/main/java/dev/relism/flash/bytes/Pairs.java

@@ -0,0 +1,42 @@
+package dev.relism.flash.bytes;
+
+/**
+ * The allocation-free idiom for returning two {@code int}s from a method without an object:
+ * pack both into one {@code long}, unpack at the call site. Already used, hand-rolled, in four
+ * places ({@code Http1HeaderMap.findFirst}, {@code QueryParams.findFirst}, and others) before this
+ * class existed — this is the single named home for the shifts so they are not duplicated (and
+ * potentially inconsistently duplicated — e.g. one copy masking with {@code 0xFFFFFFFFL} and
+ * another forgetting to) five times over.
+ *
+ * <h3>Why this works</h3>
+ * A {@code long} is 64 bits; each packed {@code int} is 32. {@link #pack} left-shifts the high
+ * half into the top 32 bits and OR's the low half into the bottom 32. {@link #lo} must mask with
+ * {@code 0xFFFFFFFFL} rather than simply cast to {@code int} after no mask, because a right-shift
+ * of a negative {@code long} sign-extends — the mask discards everything above bit 31 before the
+ * narrowing cast happens implicitly. {@link #hi} needs no mask: a right-shift by 32 already
+ * leaves only the original high bits in the low 32 positions of the result.
+ *
+ * <h3>Encoding convention used across this codebase</h3>
+ * Every {@code findFirst}-shaped method in this codebase packs {@code (start << 32) | length},
+ * i.e. {@code hi() == start} and {@code lo() == length}. {@code -1L} is the shared "not found"
+ * sentinel (a valid {@code (start, length)} pair can never be negative, since both halves are
+ * non-negative offsets/lengths).
+ */
+public final class Pairs {
+    private Pairs() {}
+
+    /** Packs two {@code int}s into one {@code long}: {@code hi} in the upper 32 bits, {@code lo} in the lower 32. */
+    public static long pack(int hi, int lo) {
+        return ((long) hi << 32) | (lo & 0xFFFFFFFFL);
+    }
+
+    /** Extracts the upper 32 bits packed by {@link #pack}. */
+    public static int hi(long packed) {
+        return (int) (packed >> 32);
+    }
+
+    /** Extracts the lower 32 bits packed by {@link #pack}. */
+    public static int lo(long packed) {
+        return (int) (packed & 0xFFFFFFFFL);
+    }
+}

flash/src/main/java/dev/relism/flash/bytes/PooledSlice.java

@@ -0,0 +1,52 @@
+package dev.relism.flash.bytes;
+
+/**
+ * per-call {@code new ByteView() { ... }} anonymous-class allocation that used to live in
+ * {@code Http1HeaderMap.view}, {@code QueryParams.view}, and {@code PathParams.view}: instead of
+ * allocating a fresh view object (plus its capturing instance) on every call, a small
+ * {@link SlicePool} of these hands out an existing instance, repositioned in place.
+ *
+ * <h3>Lifetime contract</h3>
+ * A {@code PooledSlice} handed out by {@link SlicePool#acquire} is valid only until the pool
+ * wraps around and reuses the same slot — see {@link SlicePool}'s own Javadoc for the exact
+ * "valid until the Nth subsequent acquire, or end of request" rule the owning class (e.g.
+ * {@code Http1HeaderMap}) documents precisely for its own {@code view()} method. Never retain a
+ * {@code PooledSlice} past that window, for the same reason the old anonymous view could not be
+ * retained past the handler: the bytes (and, here, additionally the slice object itself) are
+ * about to be repositioned out from under a stale reference.
+ */
+public final class PooledSlice implements ArrayBackedByteView {
+    private byte[] array;
+    private int offset;
+    private int length;
+
+    /** Repositions this slice over {@code array[offset, offset + length)}. Zero allocation. */
+    public void reset(byte[] array, int offset, int length) {
+        this.array = array;
+        this.offset = offset;
+        this.length = length;
+    }
+
+    @Override
+    public byte[] array() {
+        return array;
+    }
+
+    @Override
+    public int offset() {
+        return offset;
+    }
+
+    @Override
+    public int length() {
+        return length;
+    }
+
+    @Override
+    public byte byteAt(int index) {
+        if (index < 0 || index >= length) {
+            throw new IndexOutOfBoundsException("Index " + index + " out of bounds for length " + length);
+        }
+        return array[offset + index];
+    }
+}

flash/src/main/java/dev/relism/flash/bytes/SegmentedByteView.java

@@ -0,0 +1,80 @@
+package dev.relism.flash.bytes;
+
+import dev.relism.fpr.core.ByteView;
+
+/**
+ * A {@link ByteView} over up to {@code K} discontiguous {@code byte[]} segments, presented as one
+ * logical byte sequence. Exists for the one case in this codebase where a "single contiguous
+ * slice of one buffer" model (every other {@link ByteView} implementation) does not hold: an
+ * HPACK header block whose encoding spans more than one {@code CONTINUATION} frame (RFC 9113
+ * §6.10), where each frame's payload lives in its own connection-buffer region.
+ *
+ * <h3>Deliberately not array-backed</h3>
+ * This does not implement {@link ArrayBackedByteView} — there is no single {@code (array,
+ * offset)} pair that describes it — and {@link #supportsLong()} returns {@code false}
+ * time path is only sound for a genuinely contiguous backing array; see
+ * {@code FastPathViews.MethodPathByteView} for the other deliberately-segmented view in this
+ * codebase, which makes the same choice for the same reason).
+ *
+ * <h3>Reusable, not allocated per block</h3>
+ * {@link #reset} repositions this view over a new set of segments without allocating — the same
+ * idiom {@link PooledSlice} uses for the contiguous case. The {@code segments}/{@code offsets}/
+ * {@code lengths} arrays passed to {@link #reset} are retained by reference, not copied; the
+ * caller owns their lifetime (typically the connection's HPACK scratch, sized to
+ * {@code Http2Limits.MAX_CONTINUATION_FRAMES_PER_BLOCK}).
+ *
+ * <h3>Cost model</h3>
+ * {@link #byteAt} walks the segment table to find which segment an index falls in — O(segments),
+ * not O(1) — because this view exists precisely for the rare, deliberately-bounded case
+ * (at most {@code MAX_CONTINUATION_FRAMES_PER_BLOCK} segments); optimizing it further would add
+ * complexity for a path that, by construction, is never hot.
+ */
+public final class SegmentedByteView implements ByteView {
+    private byte[][] segments;
+    private int[] offsets;
+    private int[] lengths;
+    private int count;
+    private int totalLength;
+
+    /**
+     * Repositions this view over {@code segments[0..count)}, where segment {@code i} contributes
+     * bytes {@code segments[i][offsets[i], offsets[i] + lengths[i])}. Zero allocation: the three
+     * arrays are retained by reference.
+     */
+    public void reset(byte[][] segments, int[] offsets, int[] lengths, int count) {
+        this.segments = segments;
+        this.offsets = offsets;
+        this.lengths = lengths;
+        this.count = count;
+        int total = 0;
+        for (int i = 0; i < count; i++) total += lengths[i];
+        this.totalLength = total;
+    }
+
+    @Override
+    public int length() {
+        return totalLength;
+    }
+
+    @Override
+    public byte byteAt(int index) {
+        if (index < 0 || index >= totalLength) {
+            throw new IndexOutOfBoundsException("Index " + index + " out of bounds for length " + totalLength);
+        }
+        int remaining = index;
+        for (int i = 0; i < count; i++) {
+            int len = lengths[i];
+            if (remaining < len) {
+                return segments[i][offsets[i] + remaining];
+            }
+            remaining -= len;
+        }
+        throw new IndexOutOfBoundsException("Index " + index + " out of bounds for length " + totalLength);
+    }
+
+    /** Always {@code false} — see the class Javadoc for why a cross-segment word read is unsound. */
+    @Override
+    public boolean supportsLong() {
+        return false;
+    }
+}

flash/src/main/java/dev/relism/flash/bytes/SlicePool.java

@@ -0,0 +1,51 @@
+package dev.relism.flash.bytes;
+
+/**
+ * A small, fixed-size ring of {@link PooledSlice} instances — one per {@code ConnectionScratch}-
+ * {@code Http1HeaderMap.view}, {@code QueryParams.view}, {@code PathParams.view}).
+ *
+ * <h3>Why a ring, not a single reused slice</h3>
+ * A single reused slice (the shape {@code Http1HeaderMap.forEach} already uses for its two
+ * {@code nameSlice}/{@code valueSlice} fields) is correct only when the caller is guaranteed to
+ * finish with one slice before the next is produced — true for a single {@code forEach} callback
+ * invocation, false for {@code view()}: a handler might reasonably call
+ * {@code headers.view("A")} and {@code headers.view("B")} and want to compare both. A ring of
+ * {@code size} slices lets up to {@code size} calls' results stay simultaneously valid.
+ *
+ * <h3>Lifetime contract</h3>
+ * A slice returned by {@link #acquire} is valid until either the request ends, or {@link #acquire}
+ * is called {@code size} more times on the same pool (at which point the ring has wrapped around
+ * and repositioned that same slot for a new caller) — whichever comes first. This must be
+ * restated precisely on every method that hands out a slice from a pool (see
+ * {@code Http1HeaderMap.view}'s Javadoc for the canonical wording); it is a real, testable hazard, not
+ * a hypothetical one — see {@code SlicePoolTest#wraparoundAliasesThePreviouslyReturnedSlice} for
+ * a demonstration.
+ */
+public final class SlicePool {
+    private final PooledSlice[] slices;
+    private int next = 0;
+
+    /** A ring of {@code size} reusable slices. {@code size} must be at least 1. */
+    public SlicePool(int size) {
+        if (size < 1) throw new IllegalArgumentException("SlicePool size must be at least 1: " + size);
+        slices = new PooledSlice[size];
+        for (int i = 0; i < size; i++) slices[i] = new PooledSlice();
+    }
+
+    /** How many slices this pool cycles through before a caller's slice is reused. */
+    public int size() {
+        return slices.length;
+    }
+
+    /**
+     * Returns the next slice in the ring, repositioned over {@code array[offset, offset + length)}.
+     * Zero allocation — the returned instance already existed.
+     */
+    public PooledSlice acquire(byte[] array, int offset, int length) {
+        PooledSlice slice = slices[next];
+        next++;
+        if (next == slices.length) next = 0;
+        slice.reset(array, offset, length);
+        return slice;
+    }
+}

flash/src/main/java/dev/relism/flash/exceptions/MalformedRequestException.java

@@ -0,0 +1,22 @@
+package dev.relism.flash.exceptions;
+
+/**
+ * Thrown by the HTTP/1.1 parser when a request violates a protocol rule that must be rejected
+ * outright — most importantly the request-smuggling defenses of RFC 9112 §6.1 (see
+ *
+ * <p>Distinct from {@link HttpException}, which a <em>handler</em> throws to describe an
+ * application-level failure and which is routed through the user's configured exception
+ * handler ({@code AbstractRouter.getExceptionHandler()}). A malformed request never reaches a
+ * handler, or middleware, or the user's exception handler at all: it is rejected by the
+ * transport itself, with a fixed, minimal, non-customizable response, and the connection is
+ * always closed afterwards — never kept alive. Keeping a connection alive after a rejected
+ * request is exactly the situation a smuggling attempt exploits (a rejected first request
+ * hiding a crafted second one in the same TCP stream), so the transport never offers that
+ * choice to user code.
+ */
+public class MalformedRequestException extends HttpException {
+
+    public MalformedRequestException(int status, String message) {
+        super(status, message);
+    }
+}

flash/src/main/java/dev/relism/flash/extension/AnnotationProcessor.java

@@ -7,7 +7,7 @@ import java.util.List;
 
 /**
  * Inspects a handler class at registration time and returns zero or more
- * {@link Middleware middlewares} to inject automatically.
+ * {@link MiddlewareNode middleware nodes} to inject automatically.
  *
  * <p>Processors are called once per register call, before
  * the handler is compiled into the router. Returning an empty list is always

flash/src/main/java/dev/relism/flash/extension/FlashConfiguration.java

@@ -1,13 +1,11 @@
 package dev.relism.flash.extension;
 
 import dev.relism.flash.tls.TlsConfig;
-
+import java.util.List;
 import lombok.Builder;
 import lombok.Singular;
 import lombok.Value;
 
-import java.util.List;
-
 /**
  * Configuration for a {@link FlashApp} instance.
  *
@@ -39,27 +37,128 @@ import java.util.List;
 @Builder
 public class FlashConfiguration {
 
-    int    port;
-    String host;
+  int port;
+  String host;
 
-    /** TLS for the single default listener ({@link #port}/{@link #host}). Ignored if {@link #listeners} is non-empty. */
-    TlsConfig tls;
+  /**
+   * TLS for the single default listener ({@link #port}/{@link #host}). Ignored if {@link
+   * #listeners} is non-empty.
+   */
+  TlsConfig tls;
 
-    /** One or more listeners for this app. Non-empty list takes precedence over {@link #port}/{@link #host}/{@link #tls}. */
-    @Singular
-    List<Listener> listeners;
+  /**
+   * One or more listeners for this app. Non-empty list takes precedence over {@link #port}/{@link
+   * #host}/{@link #tls}.
+   */
+  @Singular List<Listener> listeners;
 
-    /** Maximum size of the request header buffer in bytes. Default: 64 KB. */
-    @Builder.Default
-    int maxHeaderBufferSize = 64 * 1024;
+  /** Maximum size of the request header buffer in bytes. Default: 64 KB. */
+  @Builder.Default int maxHeaderBufferSize = 64 * 1024;
 
-    /** Per-connection WebSocket read buffer size in bytes. Default: 64 KB. */
-    @Builder.Default
-    int wsFrameBufferSize = 64 * 1024;
+  /** Per-connection WebSocket read buffer size in bytes. Default: 64 KB. */
+  @Builder.Default int wsFrameBufferSize = 64 * 1024;
 
-    /** One bind target: a TCP port, an optional bind host (default: all interfaces), and optional TLS. */
-    public record Listener(int port, String host, TlsConfig tls) {
-        public Listener(int port)               { this(port, null, null); }
-        public Listener(int port, TlsConfig tls) { this(port, null, tls); }
+  /**
+   * Maximum time, in milliseconds, allowed for a request's headers to be fully read once the first
+   * byte of it has arrived. Bounds the classic slowloris attack: a peer that trickles one header
+   * byte every few seconds forever. Enforced by an absolute deadline (see {@code
+   * dev.relism.flash.transport.BufferedByteSource}), not merely a per-read socket timeout — a
+   * per-read timeout alone never trips as long as each individual read succeeds within the window,
+   * no matter how long the overall header block takes. Default: 10 000
+   */
+  @Builder.Default int headerReadTimeoutMs = 10_000;
+
+  /**
+   * Maximum time, in milliseconds, a keep-alive connection may sit idle waiting for its next
+   * request before being closed. More generous than {@link #headerReadTimeoutMs} because an idle
+   * keep-alive connection is normal, expected behaviour, not an attack in progress — the tighter
+   * bound applies only once bytes have actually started arriving. Default: 60 000
+   */
+  @Builder.Default int idleKeepAliveTimeoutMs = 60_000;
+
+  /**
+   * Maximum time, in milliseconds, a request's body may take to be fully read (by the handler or by
+   * the automatic drain after it returns) once headers are parsed. Default: 30 000
+   */
+  @Builder.Default int bodyReadTimeoutMs = 30_000;
+
+  /**
+   * Maximum time, in milliseconds, {@link dev.relism.flash.ServerHandle#stop()} waits for in-flight
+   * requests to finish after it stops accepting new connections, before force-
+   */
+  @Builder.Default int shutdownDrainTimeoutMs = 15_000;
+
+  /**
+   * Maximum connections admitted across all listeners before new connections are closed
+   * immediately at accept time, before any per-connection state (TLS handshake, protocol
+   * negotiation, HPACK tables, buffers) is set up. Defaults to an auto-scaled budget based on the
+   * JVM's max heap ({@link dev.relism.flash.transport.TransportLimits#defaultMaxConnections()}),
+   * so a connection flood cannot exhaust the heap out of the box. Set explicitly if you know your
+   * deployment's real capacity, or to {@code 0} to disable the check entirely (unlimited).
+   */
+  @Builder.Default int maxConnections =
+      dev.relism.flash.transport.TransportLimits.defaultMaxConnections();
+
+  /** Whether TLS listeners advertise HTTP/2 through ALPN. */
+  @Builder.Default boolean http2Enabled = false;
+
+  /**
+   * Whether plaintext listeners accept the HTTP/2 prior-knowledge preface. This is independent
+   * from TLS HTTP/2 and deliberately disabled by default.
+   */
+  @Builder.Default boolean http2CleartextEnabled = false;
+
+  /**
+   * Whether runtime-generated HTTP/2 header values use HPACK Huffman coding. Constants are always
+   * compressed once at startup; leaving this disabled avoids a per-byte encode pass on responses.
+   */
+  @Builder.Default boolean h2HuffmanDynamicValues = false;
+
+  /** Maximum peer RST_STREAM frames per rolling interval. */
+  @Builder.Default int h2MaxResetStreamsPerInterval =
+      dev.relism.flash.http2.Http2Limits.MAX_RESET_STREAMS_PER_INTERVAL;
+
+  /** Maximum peer-created streams per rolling interval. */
+  @Builder.Default int h2MaxStreamsCreatedPerInterval =
+      dev.relism.flash.http2.Http2Limits.MAX_STREAMS_CREATED_PER_INTERVAL;
+
+  /** Rolling interval used by HTTP/2 abuse-rate counters. */
+  @Builder.Default long h2AbuseRateIntervalMs =
+      dev.relism.flash.http2.Http2Limits.RESET_RATE_INTERVAL_MS;
+
+  /** Maximum total streams served by one HTTP/2 connection; zero disables the budget. */
+  @Builder.Default long h2MaxStreamsPerConnection =
+      dev.relism.flash.http2.Http2Limits.MAX_STREAMS_PER_CONNECTION;
+
+  /** Maximum wire bytes read by one HTTP/2 connection; zero disables the budget. */
+  @Builder.Default long h2MaxBytesPerConnection =
+      dev.relism.flash.http2.Http2Limits.MAX_BYTES_PER_CONNECTION;
+
+  /** Maximum HTTP/2 connection lifetime in milliseconds; zero disables the budget. */
+  @Builder.Default long h2MaxConnectionLifetimeMs =
+      dev.relism.flash.http2.Http2Limits.MAX_CONNECTION_LIFETIME_MS;
+
+  /** Maximum inactivity time for an open HTTP/2 stream. */
+  @Builder.Default long h2StreamIdleTimeoutMs =
+      dev.relism.flash.http2.Http2Limits.STREAM_IDLE_TIMEOUT_MS;
+
+  /**
+   * Whether every response includes a {@code Date} header (RFC 9110 §6.6.1). Default {@code true};
+   * set {@code false} if Flash sits behind a reverse proxy that already adds one, to skip the
+   * (already cheap — see {@code dev.relism.flash.http.DateHeader}) write.
+   */
+  @Builder.Default boolean sendDate = true;
+
+  /**
+   * One bind target: a TCP port, an optional bind host (default: all interfaces), and optional TLS.
+   */
+  public record Listener(int port, String host, TlsConfig tls) {
+    public Listener(int port) {
+      this(port, null, null);
     }
+
+    public Listener(int port, TlsConfig tls) {
+      this(port, null, tls);
+    }
+  }
 }

flash/src/main/java/dev/relism/flash/extension/PackageScanner.java

@@ -2,8 +2,9 @@ package dev.relism.flash.extension;
 
 import dev.relism.flash.exceptions.InitializationException;
 import dev.relism.flash.models.RequestHandler;
-import dev.relism.flash.routing.Ws;
+import dev.relism.flash.routing.Route;
 import dev.relism.flash.routing.Routes;
+import dev.relism.flash.routing.Ws;
 import dev.relism.flash.websocket.WebSocketEndpoint;
 
 import java.io.File;

flash/src/main/java/dev/relism/flash/http/ContentType.java

@@ -1,65 +1,88 @@
 package dev.relism.flash.http;
 
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http2.hpack.HpackEncoder;
+import java.nio.charset.StandardCharsets;
+import java.util.Arrays;
 import lombok.Getter;
 
-import java.nio.charset.StandardCharsets;
-
 /**
- * Pre-compiled byte representations of common HTTP {@code Content-Type} values.
- * {@link #getBytes()} returns the pre-computed array directly, never allocates.
+ * Pre-compiled byte representations of common HTTP {@code Content-Type} values. {@code getBytes()}
+ * returns the pre-computed array directly, never allocates.
  */
 @Getter
 public enum ContentType {
+  NONE(""),
 
-    NONE                (""),
+  // Text
+  TEXT_PLAIN("text/plain"),
+  TEXT_HTML("text/html"),
+  TEXT_CSS("text/css"),
+  TEXT_JAVASCRIPT("text/javascript"),
+  TEXT_XML("text/xml"),
+  TEXT_CSV("text/csv"),
+  TEXT_MARKDOWN("text/markdown"),
+  TEXT_EVENT_STREAM("text/event-stream"),
 
-    // Text
-    TEXT_PLAIN          ("text/plain"),
-    TEXT_HTML           ("text/html"),
-    TEXT_CSS            ("text/css"),
-    TEXT_JAVASCRIPT     ("text/javascript"),
-    TEXT_XML            ("text/xml"),
-    TEXT_CSV            ("text/csv"),
-    TEXT_MARKDOWN       ("text/markdown"),
-    TEXT_EVENT_STREAM   ("text/event-stream"),
+  // Application
+  JSON("application/json"),
+  XML("application/xml"),
+  BINARY("application/octet-stream"),
+  PDF("application/pdf"),
+  ZIP("application/zip"),
+  GZIP("application/gzip"),
+  FORM_URLENCODED("application/x-www-form-urlencoded"),
+  MULTIPART_FORM("multipart/form-data"),
+  GRAPHQL("application/graphql"),
+  NDJSON("application/x-ndjson"),
+  MSGPACK("application/msgpack"),
+  CBOR("application/cbor"),
+  LD_JSON("application/ld+json"),
 
-    // Application
-    JSON                ("application/json"),
-    XML                 ("application/xml"),
-    BINARY              ("application/octet-stream"),
-    PDF                 ("application/pdf"),
-    ZIP                 ("application/zip"),
-    GZIP                ("application/gzip"),
-    FORM_URLENCODED     ("application/x-www-form-urlencoded"),
-    MULTIPART_FORM      ("multipart/form-data"),
-    GRAPHQL             ("application/graphql"),
-    NDJSON              ("application/x-ndjson"),
-    MSGPACK             ("application/msgpack"),
-    CBOR                ("application/cbor"),
-    LD_JSON             ("application/ld+json"),
+  // Image
+  IMAGE_PNG("image/png"),
+  IMAGE_JPEG("image/jpeg"),
+  IMAGE_GIF("image/gif"),
+  IMAGE_WEBP("image/webp"),
+  IMAGE_SVG("image/svg+xml"),
+  IMAGE_ICO("image/x-icon"),
+  IMAGE_AVIF("image/avif"),
 
-    // Image
-    IMAGE_PNG           ("image/png"),
-    IMAGE_JPEG          ("image/jpeg"),
-    IMAGE_GIF           ("image/gif"),
-    IMAGE_WEBP          ("image/webp"),
-    IMAGE_SVG           ("image/svg+xml"),
-    IMAGE_ICO           ("image/x-icon"),
-    IMAGE_AVIF          ("image/avif"),
+  // Font
+  FONT_WOFF("font/woff"),
+  FONT_WOFF2("font/woff2"),
 
-    // Font
-    FONT_WOFF           ("font/woff"),
-    FONT_WOFF2          ("font/woff2"),
+  // Audio / Video
+  AUDIO_MPEG("audio/mpeg"),
+  AUDIO_OGG("audio/ogg"),
+  VIDEO_MP4("video/mp4"),
+  VIDEO_WEBM("video/webm");
 
-    // Audio / Video
-    AUDIO_MPEG          ("audio/mpeg"),
-    AUDIO_OGG           ("audio/ogg"),
-    VIDEO_MP4           ("video/mp4"),
-    VIDEO_WEBM          ("video/webm");
+  private final byte[] bytes;
+  private final byte[] hpackBytes;
+  private static final ContentType[] ALL = values();
 
-    private final byte[] bytes;
-
-    ContentType(String value) {
-        this.bytes = value.getBytes(StandardCharsets.UTF_8);
+  ContentType(String value) {
+    this.bytes = value.getBytes(StandardCharsets.UTF_8);
+    if (bytes.length == 0) {
+      this.hpackBytes = bytes;
+    } else {
+      ByteWriter out = new ByteWriter(32);
+      HpackEncoder.writeLiteralWithNameIndex(out, 31, bytes, true);
+      this.hpackBytes = Arrays.copyOf(out.array(), out.length());
     }
+  }
+
+  /** Precompiled HPACK {@code content-type} field, or an empty array for {@link #NONE}. */
+  public byte[] getHpackBytes() {
+    return hpackBytes;
+  }
+
+  /** Finds the boot-time HPACK rendering for a response content-type byte array. */
+  public static byte[] hpackBytesFor(byte[] value) {
+    for (ContentType type : ALL) {
+      if (type.bytes == value || Arrays.equals(type.bytes, value)) return type.hpackBytes;
+    }
+    return null;
+  }
 }

flash/src/main/java/dev/relism/flash/http/DateHeader.java

@@ -0,0 +1,77 @@
+package dev.relism.flash.http;
+
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http2.hpack.HpackEncoder;
+import java.nio.charset.StandardCharsets;
+import java.time.ZoneOffset;
+import java.time.ZonedDateTime;
+import java.time.format.DateTimeFormatter;
+import java.util.Arrays;
+import java.util.Locale;
+
+/**
+ * A daemon refreshes both protocol renderings once per second. Response writers only perform one
+ * volatile read and copy already-encoded bytes into their output buffer.
+ */
+public final class DateHeader {
+
+  private DateHeader() {}
+
+  private static final DateTimeFormatter FORMATTER =
+      DateTimeFormatter.ofPattern("EEE, dd MMM yyyy HH:mm:ss 'GMT'", Locale.US)
+          .withZone(ZoneOffset.UTC);
+
+  private record Snapshot(byte[] http1, byte[] hpack) {}
+
+  private static volatile Snapshot current = encode();
+
+  static {
+    Thread refresher =
+        new Thread(
+            () -> {
+              while (true) {
+                try {
+                  Thread.sleep(1000);
+                } catch (InterruptedException e) {
+                  Thread.currentThread().interrupt();
+                  return;
+                }
+                current = encode();
+              }
+            },
+            "flash-date-header");
+    refresher.setDaemon(true);
+    refresher.start();
+  }
+
+  private static Snapshot encode() {
+    byte[] value = format(ZonedDateTime.now(ZoneOffset.UTC)).getBytes(StandardCharsets.US_ASCII);
+    byte[] prefix = "Date: ".getBytes(StandardCharsets.US_ASCII);
+    byte[] http1 = new byte[prefix.length + value.length + 2];
+    System.arraycopy(prefix, 0, http1, 0, prefix.length);
+    System.arraycopy(value, 0, http1, prefix.length, value.length);
+    http1[http1.length - 2] = '\r';
+    http1[http1.length - 1] = '\n';
+
+    ByteWriter encoded = new ByteWriter(32);
+    HpackEncoder.writeLiteralWithNameIndex(encoded, 33, value, true);
+    return new Snapshot(http1, Arrays.copyOf(encoded.array(), encoded.length()));
+  }
+
+  static String format(ZonedDateTime time) {
+    return FORMATTER.format(time);
+  }
+
+  /**
+   * The current pre-encoded {@code "Date: ...\r\n"} field line, accurate to within one second.
+   * Never allocates — the same array is returned until the next refresh.
+   */
+  public static byte[] bytes() {
+    return current.http1;
+  }
+
+  /** Current precompiled HPACK {@code date} field. */
+  public static byte[] hpackBytes() {
+    return current.hpack;
+  }
+}

flash/src/main/java/dev/relism/flash/http/HopByHopHeaders.java

@@ -0,0 +1,87 @@
+package dev.relism.flash.http;
+
+import dev.relism.flash.models.HeaderView;
+import dev.relism.fpr.core.ByteView;
+
+/** Shared proxy policy for fields that must not cross an HTTP connection boundary. */
+public final class HopByHopHeaders {
+  public enum Protocol {
+    HTTP_1_1,
+    HTTP_2
+  }
+
+  private HopByHopHeaders() {}
+
+  /** Returns whether a field may be copied to a new downstream connection. */
+  public static boolean shouldForward(
+      HeaderView source,
+      ByteView name,
+      ByteView value,
+      Protocol sourceProtocol,
+      Protocol targetProtocol) {
+    if (name.length() == 0 || name.byteAt(0) == ':') return false;
+    if (is(name, "connection")
+        || is(name, "keep-alive")
+        || is(name, "proxy-connection")
+        || is(name, "proxy-authenticate")
+        || is(name, "proxy-authorization")
+        || is(name, "trailer")
+        || is(name, "transfer-encoding")
+        || is(name, "upgrade")) {
+      return false;
+    }
+    if (isConnectionListed(source, name)) return false;
+    if (is(name, "te")) {
+      return targetProtocol == Protocol.HTTP_2 && isTrimmed(value, "trailers");
+    }
+    return true;
+  }
+
+  private static boolean isConnectionListed(HeaderView source, ByteView fieldName) {
+    for (String value : source.all("connection")) {
+      int start = 0;
+      while (start < value.length()) {
+        int comma = value.indexOf(',', start);
+        int end = comma < 0 ? value.length() : comma;
+        while (start < end && isWhitespace(value.charAt(start))) start++;
+        while (end > start && isWhitespace(value.charAt(end - 1))) end--;
+        if (equalsAsciiIgnoreCase(fieldName, value, start, end)) return true;
+        start = comma < 0 ? value.length() : comma + 1;
+      }
+    }
+    return false;
+  }
+
+  private static boolean is(ByteView bytes, String expected) {
+    return equalsAsciiIgnoreCase(bytes, expected, 0, expected.length());
+  }
+
+  private static boolean isTrimmed(ByteView bytes, String expected) {
+    int start = 0;
+    int end = bytes.length();
+    while (start < end && isWhitespace((char) bytes.byteAt(start))) start++;
+    while (end > start && isWhitespace((char) bytes.byteAt(end - 1))) end--;
+    if (end - start != expected.length()) return false;
+    for (int i = 0; i < expected.length(); i++) {
+      if (lower(bytes.byteAt(start + i) & 0xff) != lower(expected.charAt(i))) return false;
+    }
+    return true;
+  }
+
+  private static boolean equalsAsciiIgnoreCase(
+      ByteView bytes, String expected, int expectedStart, int expectedEnd) {
+    if (bytes.length() != expectedEnd - expectedStart) return false;
+    for (int i = 0; i < bytes.length(); i++) {
+      if (lower(bytes.byteAt(i) & 0xff) != lower(expected.charAt(expectedStart + i))) return false;
+    }
+    return true;
+  }
+
+  private static int lower(int value) {
+    return value >= 'A' && value <= 'Z' ? value + ('a' - 'A') : value;
+  }
+
+  private static boolean isWhitespace(char value) {
+    return value == ' ' || value == '\t';
+  }
+}

flash/src/main/java/dev/relism/flash/http/Http1Limits.java

@@ -0,0 +1,156 @@
+package dev.relism.flash.http;
+
+/**
+ * Bounds the HTTP/1.1 parser ({@code RequestParser}, {@code ChunkedInputStream}) enforces
+ * against a peer's input, in one place.
+ *
+ * against a named constant here — never against an ad-hoc literal, and never by letting the
+ * underlying buffer throw on overrun. Each field's Javadoc names the specific attack it bounds.
+ *
+ * Compare {@code dev.relism.flash.http2.Http2Limits}, the HTTP/2 equivalent.
+ */
+public final class Http1Limits {
+
+    private Http1Limits() {
+    }
+
+    /**
+     * The largest {@code Content-Length} value accepted, in bytes. RFC 9112 places no upper
+     * bound on the header's numeric value, but an unbounded value from a hostile peer is a
+     * resource-exhaustion vector for any code path that pre-sizes a buffer from it. Requests
+     * declaring a length above this are rejected with {@code 413 Payload Too Large} before any
+     * body byte is read.
+     *
+     * <p>4 GiB — generous enough for legitimate large uploads (Flash is a general-purpose
+     * server, not an API-only framework with a tiny default), while still bounding a hostile
+     * peer to a finite, known-in-advance number rather than the effectively unbounded
+     * {@code Long.MAX_VALUE} the parser accepted before this limit existed. Comfortably above
+     * {@code Integer.MAX_VALUE} (~2.1 billion) so legitimate very-large declared lengths are
+     */
+    public static final long MAX_CONTENT_LENGTH = 4L * 1024 * 1024 * 1024;
+
+    /**
+     * Maximum number of header lines accepted in a single request. Without this bound, a
+     * request with tens of thousands of one-byte headers passes the total header-block size
+     * check ({@code maxHeaderBufferSize}) while still forcing every subsequent
+     * {@code Http1HeaderMap} lookup to scan all of them — turning a small request into quadratic CPU
+     */
+    public static final int MAX_HEADER_COUNT = 100;
+
+    /**
+     * Maximum length, in bytes, of a single header field name. RFC 9110 §5.1 places no formal
+     * limit; this bound exists purely to cap per-header memory and scan cost.
+     */
+    public static final int MAX_HEADER_NAME_LENGTH = 256;
+
+    /**
+     * Maximum length, in bytes, of a single header field value. Bounds per-header memory and
+     * scan cost the same way {@link #MAX_HEADER_NAME_LENGTH} bounds the name.
+     */
+    public static final int MAX_HEADER_VALUE_LENGTH = 8_192;
+
+    /**
+     * Maximum length, in bytes, of the request line ({@code METHOD SP target SP version}).
+     * Tracked separately from the overall header-buffer size so an oversized request line is
+     * rejected with a specific, correct status ({@code 414 URI Too Long}) rather than folded
+     * into the generic header-block-too-large case.
+     */
+    public static final int MAX_REQUEST_LINE_LENGTH = 8_192;
+
+    /**
+     * Maximum size, in bytes, of a single {@code Transfer-Encoding: chunked} chunk.
+     * {@code ChunkedInputStream.readChunkSize} previously accepted any value up to 2 GiB before
+     * rejecting it; a hostile peer can advertise a huge chunk size and then trickle bytes,
+     * forcing the connection to stay open far longer than any legitimate chunk would need
+     * (bounded separately by {@code bodyReadTimeoutMs}, but this limit catches the size claim
+     * itself before that timeout would).
+     */
+    public static final long MAX_CHUNK_SIZE = 16L * 1024 * 1024;
+
+    /**
+     * Maximum length, in bytes, of the chunk-extension section (the optional
+     * {@code ;name=value} data after a chunk size and before its CRLF, RFC 9112 §7.1.1). Flash
+     * does not interpret chunk extensions; without a bound, a peer could send an arbitrarily
+     * long extension on every chunk purely to waste CPU discarding it.
+     */
+    public static final int MAX_CHUNK_EXT_LENGTH = 256;
+
+    /**
+     * Maximum number of chunks accepted in a single request body. Without this bound, a peer
+     * can send an unbounded number of minimal (or zero-length) chunks, each cheap individually
+     * but collectively forcing unbounded per-chunk framing work — a "death by a thousand
+     * chunks" variant of a slow-body attack.
+     */
+    public static final int MAX_CHUNKS_PER_BODY = 100_000;
+
+    /**
+     * Maximum number of trailer header lines accepted after the final chunk of a chunked body
+     * (RFC 9112 §7.1.2). Bounded for the same reason {@link #MAX_HEADER_COUNT} bounds the
+     * regular header section; trailer values are separately bounded by
+     * {@link #MAX_HEADER_VALUE_LENGTH}.
+     */
+    public static final int MAX_TRAILER_COUNT = 50;
+
+    /**
+     * buffer as the response head (status line + headers) and written with it in a single
+     * {@code OutputStream.write} call; larger bodies are written in a second {@code write} right
+     * after the head, since copying a large body into the head buffer first would cost more
+     * (an extra full-body memcpy) than the syscall it saves. 8 KiB — matches this codebase's
+     * other "one socket-buffer's worth" constants ({@code ConnectionScratch.RELAY_BUFFER_SIZE},
+     * {@code BufferedByteSource.DEFAULT_BUFFER_SIZE}) rather than introducing an uncalibrated
+     */
+    public static final int INLINE_BODY_THRESHOLD = 8192;
+
+    /**
+     * {@code multipart/form-data} body. Without this bound, a peer can send an unbounded number
+     * of minimal parts — each cheap individually but forcing unbounded growth of the parser's
+     * {@code scanned} list and unbounded per-part header-parsing work, the multipart analogue of
+     * {@link #MAX_CHUNKS_PER_BODY}.
+     */
+    public static final int MAX_MULTIPART_PARTS = 1_000;
+
+    /**
+     * {@code Content-Type}, …) accepted per multipart part. Real clients send at most two or
+     * three; without a bound a peer could send an effectively unlimited number before the blank
+     * line that ends a part's header block, forcing unbounded {@code HashMap} growth per part.
+     */
+    public static final int MAX_MULTIPART_PART_HEADER_COUNT = 20;
+
+    /**
+     * header block. {@code Multipart.readLine} otherwise has no bound of its own to fall back
+     * on — unlike the top-level HTTP headers (bounded by {@link #MAX_HEADER_VALUE_LENGTH} in
+     * {@code RequestParser}), a line here with no {@code \r\n} would grow its {@code StringBuilder}
+     * without limit for as long as the peer keeps streaming bytes.
+     */
+    public static final int MAX_MULTIPART_HEADER_LINE_LENGTH = 8_192;
+
+    /**
+     * buffers eagerly into a {@code byte[]} — text fields (always buffered) and, during a full
+     * {@code parts()}/{@code parts(String)} scan, file bodies too. {@link #MAX_CONTENT_LENGTH}
+     * bounds the whole request body, but at 4 GiB (and effectively unbounded for a chunked body,
+     * see {@link #MAX_CHUNKS_PER_BODY} × {@link #MAX_CHUNK_SIZE}) it does nothing to stop a
+     * single part from exhausting the heap on its own — this is the bound that actually protects
+     * {@code ByteArrayOutputStream}-style eager buffering. Deliberately does not apply to
+     * {@code Part.materialize()} on a streaming file part returned by {@code Multipart.file()} —
+     * that call is documented as an explicit, opt-in heap allocation the caller chooses to pay for.
+     */
+    public static final long MAX_MULTIPART_BUFFERED_PART_SIZE = 10L * 1024 * 1024;
+
+    /**
+     * Maximum combined size, in bytes, of every response header's name + value bytes
+     * ({@code Response.header(...)}'s growable {@code headerRegion}). Unlike every other bound in
+     * this class, this one guards against a bug in <em>Flash's own caller</em> rather than a
+     * hostile peer — a handler that calls {@code header(...)} in an unbounded loop (e.g. echoing
+     * an unbounded collection into headers) would otherwise grow this connection's scratch region
+     * without limit for the rest of its lifetime, since it is never shrunk back down between
+     */
+    public static final int MAX_RESPONSE_HEADER_BYTES = 65_536;
+
+    /**
+     * Maximum number of {@code Response.header(...)} calls (any overload) accepted on a single
+     * response. Same rationale as {@link #MAX_RESPONSE_HEADER_BYTES}: bounds the response-side
+     * analogue of {@link #MAX_HEADER_COUNT}, since an unbounded call count grows the header index
+     * arrays even if each individual header is small.
+     */
+    public static final int MAX_RESPONSE_HEADER_COUNT = 1_000;
+}

flash/src/main/java/dev/relism/flash/http/HttpStatus.java

@@ -1,101 +1,164 @@
 package dev.relism.flash.http;
 
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http2.hpack.HpackEncoder;
 import java.nio.charset.StandardCharsets;
-import java.util.List;
 
 /**
- * Pre-compiled byte representations of standard HTTP status lines.
- * Uses a direct-access array for O(1) lookup with zero allocation.
+ * Pre-compiled byte representations of standard HTTP status lines. Uses a direct-access array for
+ * O(1) lookup with zero allocation.
  */
 public enum HttpStatus {
 
-    // 1xx
-    CONTINUE                        (100, "Continue"),
-    SWITCHING_PROTOCOLS             (101, "Switching Protocols"),
+  // 1xx
+  CONTINUE(100, "Continue"),
+  SWITCHING_PROTOCOLS(101, "Switching Protocols"),
 
-    // 2xx
-    OK                              (200, "OK"),
-    CREATED                         (201, "Created"),
-    ACCEPTED                        (202, "Accepted"),
-    NO_CONTENT                      (204, "No Content"),
-    PARTIAL_CONTENT                 (206, "Partial Content"),
+  // 2xx
+  OK(200, "OK"),
+  CREATED(201, "Created"),
+  ACCEPTED(202, "Accepted"),
+  NO_CONTENT(204, "No Content"),
+  PARTIAL_CONTENT(206, "Partial Content"),
 
-    // 3xx
-    MOVED_PERMANENTLY               (301, "Moved Permanently"),
-    FOUND                           (302, "Found"),
-    NOT_MODIFIED                    (304, "Not Modified"),
-    TEMPORARY_REDIRECT              (307, "Temporary Redirect"),
-    PERMANENT_REDIRECT              (308, "Permanent Redirect"),
+  // 3xx
+  MOVED_PERMANENTLY(301, "Moved Permanently"),
+  FOUND(302, "Found"),
+  NOT_MODIFIED(304, "Not Modified"),
+  TEMPORARY_REDIRECT(307, "Temporary Redirect"),
+  PERMANENT_REDIRECT(308, "Permanent Redirect"),
 
-    // 4xx
-    BAD_REQUEST                     (400, "Bad Request"),
-    UNAUTHORIZED                    (401, "Unauthorized"),
-    FORBIDDEN                       (403, "Forbidden"),
-    NOT_FOUND                       (404, "Not Found"),
-    METHOD_NOT_ALLOWED              (405, "Method Not Allowed"),
-    NOT_ACCEPTABLE                  (406, "Not Acceptable"),
-    CONFLICT                        (409, "Conflict"),
-    GONE                            (410, "Gone"),
-    LENGTH_REQUIRED                 (411, "Length Required"),
-    PAYLOAD_TOO_LARGE               (413, "Payload Too Large"),
-    URI_TOO_LONG                    (414, "URI Too Long"),
-    UNSUPPORTED_MEDIA_TYPE          (415, "Unsupported Media Type"),
-    UNPROCESSABLE_ENTITY            (422, "Unprocessable Entity"),
-    TOO_MANY_REQUESTS               (429, "Too Many Requests"),
+  // 4xx
+  BAD_REQUEST(400, "Bad Request"),
+  UNAUTHORIZED(401, "Unauthorized"),
+  FORBIDDEN(403, "Forbidden"),
+  NOT_FOUND(404, "Not Found"),
+  METHOD_NOT_ALLOWED(405, "Method Not Allowed"),
+  NOT_ACCEPTABLE(406, "Not Acceptable"),
+  CONFLICT(409, "Conflict"),
+  GONE(410, "Gone"),
+  LENGTH_REQUIRED(411, "Length Required"),
+  PRECONDITION_FAILED(412, "Precondition Failed"),
+  PAYLOAD_TOO_LARGE(413, "Payload Too Large"),
+  URI_TOO_LONG(414, "URI Too Long"),
+  UNSUPPORTED_MEDIA_TYPE(415, "Unsupported Media Type"),
+  RANGE_NOT_SATISFIABLE(416, "Range Not Satisfiable"),
+  EXPECTATION_FAILED(417, "Expectation Failed"),
+  MISDIRECTED_REQUEST(421, "Misdirected Request"),
+  UNPROCESSABLE_ENTITY(422, "Unprocessable Entity"),
+  TOO_MANY_REQUESTS(429, "Too Many Requests"),
+  REQUEST_HEADER_FIELDS_TOO_LARGE(431, "Request Header Fields Too Large"),
 
-    // 5xx
-    INTERNAL_SERVER_ERROR           (500, "Internal Server Error"),
-    NOT_IMPLEMENTED                 (501, "Not Implemented"),
-    BAD_GATEWAY                     (502, "Bad Gateway"),
-    SERVICE_UNAVAILABLE             (503, "Service Unavailable"),
-    GATEWAY_TIMEOUT                 (504, "Gateway Timeout");
+  // 5xx
+  INTERNAL_SERVER_ERROR(500, "Internal Server Error"),
+  NOT_IMPLEMENTED(501, "Not Implemented"),
+  BAD_GATEWAY(502, "Bad Gateway"),
+  SERVICE_UNAVAILABLE(503, "Service Unavailable"),
+  GATEWAY_TIMEOUT(504, "Gateway Timeout"),
+  HTTP_VERSION_NOT_SUPPORTED(505, "HTTP Version Not Supported"),
+  INSUFFICIENT_STORAGE(507, "Insufficient Storage"),
+  NETWORK_AUTHENTICATION_REQUIRED(511, "Network Authentication Required");
 
-    private static final int MAX_STATUS_CODE = 504;
-    private static final byte[][] INDEX = new byte[MAX_STATUS_CODE + 1][];
-    private static final String[] REASONS = new String[MAX_STATUS_CODE + 1];
+  // ArrayIndexOutOfBoundsException from this static initializer the moment any constant
+  // above it (421, 431, 505, 507, 511 — several of which HTTP/2 needs, see MISDIRECTED_REQUEST
+  // and REQUEST_HEADER_FIELDS_TOO_LARGE above) was added. Computed from values() instead, so
+  // adding a status code can never silently break class loading again.
+  private static final int MAX_STATUS_CODE;
+  private static final byte[][] INDEX;
+  private static final byte[][] HPACK_INDEX;
+  private static final String[] REASONS;
 
-    static {
-        for (HttpStatus s : values()) {
-            INDEX[s.code] = s.bytes;
-            REASONS[s.code] = s.reason;
-        }
+  static {
+    int max = 0;
+    for (HttpStatus s : values()) max = Math.max(max, s.code);
+    MAX_STATUS_CODE = max;
+    INDEX = new byte[MAX_STATUS_CODE + 1][];
+    HPACK_INDEX = new byte[MAX_STATUS_CODE + 1][];
+    REASONS = new String[MAX_STATUS_CODE + 1];
+    for (HttpStatus s : values()) {
+      INDEX[s.code] = s.bytes;
+      HPACK_INDEX[s.code] = s.hpackBytes;
+      REASONS[s.code] = s.reason;
     }
+  }
 
-    private final int    code;
-    private final String reason;
-    private final byte[] bytes;
+  private final int code;
+  private final String reason;
+  private final byte[] bytes;
+  private final byte[] hpackBytes;
 
-    HttpStatus(int code, String reason) {
-        this.code  = code;
-        this.reason = reason;
-        this.bytes = (code + " " + reason).getBytes(StandardCharsets.UTF_8);
+  HttpStatus(int code, String reason) {
+    this.code = code;
+    this.reason = reason;
+    this.bytes = (code + " " + reason).getBytes(StandardCharsets.UTF_8);
+    this.hpackBytes = encodeHpack(code);
+  }
+
+  /** Numeric status code (e.g. {@code 200}). */
+  public int code() {
+    return code;
+  }
+
+  /** Pre-encoded {@code "200 OK"} bytes — zero allocation on the write path. */
+  public byte[] bytes() {
+    return bytes;
+  }
+
+  /** Precompiled HPACK representation of {@code :status}. */
+  public byte[] hpackBytes() {
+    return hpackBytes;
+  }
+
+  /** Reason phrase (e.g. {@code "OK"}). */
+  public String reason() {
+    return reason;
+  }
+
+  /**
+   * Returns pre-compiled status bytes for the given code. Access is O(1) and generates zero
+   * garbage.
+   */
+  public static byte[] bytesForCode(int code) {
+    if (code >= 0 && code <= MAX_STATUS_CODE) {
+      return INDEX[code];
     }
+    return null;
+  }
 
-    /** Numeric status code (e.g. {@code 200}). */
-    public int code() { return code; }
-
-    /** Pre-encoded {@code "200 OK"} bytes — zero allocation on the write path. */
-    public byte[] bytes() { return bytes; }
-
-    /** Reason phrase (e.g. {@code "OK"}). */
-    public String reason() { return reason; }
-
-    /**
-     * Returns pre-compiled status bytes for the given code.
-     * Access is O(1) and generates zero garbage.
-     */
-    public static byte[] bytesForCode(int code) {
-        if (code >= 0 && code <= MAX_STATUS_CODE) {
-            return INDEX[code];
-        }
-        return null;
+  /** Returns reason phrase for the given code, or null if unknown. */
+  public static String reasonForCode(int code) {
+    if (code >= 0 && code <= MAX_STATUS_CODE) {
+      return REASONS[code];
     }
+    return null;
+  }
 
-    /** Returns reason phrase for the given code, or null if unknown. */
-    public static String reasonForCode(int code) {
-        if (code >= 0 && code <= MAX_STATUS_CODE) {
-            return REASONS[code];
-        }
-        return null;
+  /** Returns the precompiled HPACK status field for a known code, or {@code null}. */
+  public static byte[] hpackBytesForCode(int code) {
+    return code >= 0 && code <= MAX_STATUS_CODE ? HPACK_INDEX[code] : null;
+  }
+
+  private static byte[] encodeHpack(int code) {
+    int staticIndex =
+        switch (code) {
+          case 200 -> 8;
+          case 204 -> 9;
+          case 206 -> 10;
+          case 304 -> 11;
+          case 400 -> 12;
+          case 404 -> 13;
+          case 500 -> 14;
+          default -> 0;
+        };
+    ByteWriter out = new ByteWriter(8);
+    if (staticIndex != 0) {
+      HpackEncoder.writeIndexed(out, staticIndex);
+    } else {
+      byte[] value = {
+        (byte) ('0' + code / 100), (byte) ('0' + code / 10 % 10), (byte) ('0' + code % 10)
+      };
+      HpackEncoder.writeLiteralWithNameIndex(out, 8, value, true);
     }
+    return java.util.Arrays.copyOf(out.array(), out.length());
+  }
 }

flash/src/main/java/dev/relism/flash/http1/Http1Connection.java

@@ -0,0 +1,141 @@
+package dev.relism.flash.http1;
+
+import dev.relism.flash.RequestParser;
+import dev.relism.flash.exceptions.MalformedRequestException;
+import dev.relism.flash.http.ContentType;
+import dev.relism.flash.http.HttpMethod;
+import dev.relism.flash.models.Request;
+import dev.relism.flash.models.RequestHandler;
+import dev.relism.flash.models.Response;
+import dev.relism.flash.transport.BufferedByteSource;
+import dev.relism.flash.transport.ConnectionContext;
+import dev.relism.flash.transport.ConnectionProtocol;
+import dev.relism.flash.websocket.WebSocketHandler;
+import dev.relism.flash.websocket.WebSocketLoop;
+import dev.relism.flash.websocket.WebSocketSession;
+import dev.relism.flash.websocket.WebSocketUpgrade;
+
+import java.io.IOException;
+import java.io.OutputStream;
+import java.net.SocketTimeoutException;
+
+/**
+ * The HTTP/1.1 keep-alive request loop: parse → route → handle → respond, repeated until the
+ * connection closes. Sole responsibility: drive that loop for one connection; parsing lives in
+ * {@link RequestParser}, serialization in {@link Http1ResponseWriter}, and the WebSocket upgrade
+ * path hands off to {@link WebSocketUpgrade}/{@link WebSocketLoop} entirely — once a connection
+ * upgrades, this class has nothing further to do with it.
+ */
+public final class Http1Connection implements ConnectionProtocol {
+
+    @Override
+    public void run(ConnectionContext ctx) throws IOException {
+        RequestParser parser = new RequestParser(
+                ctx.configuration().getMaxHeaderBufferSize(),
+                ctx.remoteAddress(),
+                ctx.sslSocket());
+
+        BufferedByteSource in  = ctx.in();
+        OutputStream        out = ctx.out();
+        byte[] idleProbe = new byte[1];
+
+        // reused across every request on this connection — see AbstractRouter#newScratch.
+        Object routeScratch   = ctx.router().newScratch();
+        Object wsRouteScratch = ctx.wsRouter().newScratch();
+
+        Response pooledResponse = new Response(200, ContentType.TEXT_PLAIN);
+
+        while (!ctx.stopped().getAsBoolean()) {
+            // idle-keep-alive timeout — sitting idle between keep-alive requests is normal, not
+            // an attack. Skipped when the parser already has bytes buffered from a previous
+            // read (HTTP pipelining): the next request has, by definition, already started, so
+            // waiting on the *source* for a fresh byte would wait for something that already
+            // arrived and is sitting in the parser's own buffer.
+            if (!parser.hasBufferedBytes()) {
+                in.setDeadline(System.nanoTime() + ctx.configuration().getIdleKeepAliveTimeoutMs() * 1_000_000L);
+                int firstByteSeen;
+                try {
+                    firstByteSeen = in.peek(idleProbe, 0, 1);
+                } catch (SocketTimeoutException e) {
+                    break; // idle timeout — nothing pending; close quietly, like EOF
+                }
+                if (firstByteSeen <= 0) break; // clean EOF
+            }
+
+            // Bytes have started arriving: tighten to the slowloris-specific bound for the rest
+            // of the header block.
+            in.setDeadline(System.nanoTime() + ctx.configuration().getHeaderReadTimeoutMs() * 1_000_000L);
+            Request request;
+            try {
+                request = parser.parse(in);
+            } catch (MalformedRequestException e) {
+                // through a handler or the user's exception handler — and the connection is
+                // always closed afterwards, never kept alive.
+                Response rejection = new Response(e.status(), e.getMessage(), ContentType.TEXT_PLAIN);
+                Http1ResponseWriter.writeResponse(out, rejection, null, false, ctx.configuration().isSendDate(), ctx.scratch());
+                break;
+            } catch (SocketTimeoutException e) {
+                break; // header-read deadline exceeded — close
+            }
+            if (request == null) break;
+
+            if (request.method() == HttpMethod.GET && WebSocketUpgrade.isWebSocketUpgrade(request)) {
+                in.clearDeadline(); // the WS session loop is long-lived; it paces itself
+                WebSocketHandler wsHandler = ctx.wsRouter().route(request, wsRouteScratch);
+                if (wsHandler == null) {
+                    out.write(WebSocketUpgrade.REJECT_400);
+                    out.flush();
+                    break;
+                }
+                // Flush buffered HTTP bytes (the 101 response) before WebSocketSession takes
+                // over rawOut — otherwise the handshake reply stays stuck in the buffered
+                // stream and the client never sees it.
+                WebSocketUpgrade.performHandshake(out, request, ctx.scratch());
+                out.flush();
+                request.drain();
+                WebSocketSession session = new WebSocketSession(
+                        in, ctx.rawOut(), ctx.configuration().getWsFrameBufferSize(), request, false);
+                WebSocketLoop.run(session, wsHandler);
+                return;
+            }
+
+            // Headers are fully read; the body (if any) may still be pending — whether the
+            // handler consumes it or the automatic drain() below does, bound it by the same
+            // deadline.
+            in.setDeadline(System.nanoTime() + ctx.configuration().getBodyReadTimeoutMs() * 1_000_000L);
+
+            boolean keepAlive = Http1KeepAlive.isKeepAlive(request);
+            Response response = pooledResponse.reset(200, ContentType.TEXT_PLAIN);
+
+            RequestHandler handler = ctx.router().route(request, routeScratch);
+            if (handler == null) handler = ctx.router().getNotFoundHandler();
+
+            try {
+                Object result = handler.handle(request, response);
+                if (result instanceof Response r) response = r;
+                else if (result != null) response.setBody(result);
+            } catch (Exception ex) {
+                Object result = ctx.router().getExceptionHandler().handle(ex, request, response);
+                if (result instanceof Response r) response = r;
+                else if (result != null) response.setBody(result);
+            }
+
+            // this handler was running (the common case: draining connections mid-request) must
+            // still force this response to Connection: close, not whatever was decided before
+            // the handler ran.
+            boolean actuallyKeepAlive = keepAlive && !ctx.stopped().getAsBoolean();
+            Http1ResponseWriter.writeResponse(out, response, request.method(), actuallyKeepAlive,
+                    ctx.configuration().isSendDate(), ctx.scratch());
+            request.drain();
+            // dropped, if the connection closes) — poison them in dev mode so any reference the
+            // handler improperly retained (a captured field, an async callback) fails loudly on
+            // its next access instead of silently reading whatever comes next. Only the pooled
+            // Response is recycled: if the handler returned a different instance, that object was
+            // never pooled in the first place and owes nothing back to this connection.
+            request.recycle();
+            if (response == pooledResponse) pooledResponse.recycle();
+            in.clearDeadline();
+            if (!actuallyKeepAlive) break;
+        }
+    }
+}

flash/src/main/java/dev/relism/flash/http1/Http1KeepAlive.java

@@ -0,0 +1,70 @@
+package dev.relism.flash.http1;
+
+import dev.relism.flash.models.Request;
+import dev.relism.fpr.core.ByteView;
+
+/**
+ * HTTP/1.1 keep-alive decision (RFC 9110 §7.6.1) and the shared {@code Connection} header
+ * token-list scanner both it and WebSocket upgrade detection need.
+ *
+ * (e.g. {@code "Connection: keep-alive, Upgrade"}), not a single value — a whole-value compare
+ * against {@code "close"} misses exactly that case. {@link #tokenListContains} is the one
+ * scanner both this class's {@link #isKeepAlive} and {@code WebSocketUpgrade}'s
+ * {@code Connection: Upgrade} check use, so the two can never drift apart again.
+ */
+public final class Http1KeepAlive {
+
+    private Http1KeepAlive() {
+    }
+
+    /**
+     * Whether the connection should remain open after this response. HTTP/1.1 defaults to
+     * keep-alive unless {@code Connection} lists {@code close}; HTTP/1.0 defaults to close
+     * unless it lists {@code keep-alive}.
+     */
+    public static boolean isKeepAlive(Request request) {
+        if (connectionContainsToken(request, "close")) return false;
+        ByteView protocol = request.getRequestLine().getProtocol();
+        int plen = protocol.length();
+        if (plen == 8) {
+            byte minor = protocol.byteAt(7);
+            if (minor == '1') return true;
+            if (minor == '0') return connectionContainsToken(request, "keep-alive");
+        }
+        return false;
+    }
+
+    /** Whether the request's {@code Connection} header lists {@code token} (case-insensitive). */
+    public static boolean connectionContainsToken(Request request, String token) {
+        ByteView conn = request.getRequestLine().getHeaders().view("Connection");
+        if (conn == null) return false;
+        return tokenListContains(conn, token);
+    }
+
+    /** Scans a comma-separated token list for {@code token} (case-insensitive, OWS-tolerant). */
+    public static boolean tokenListContains(ByteView view, String token) {
+        int len = view.length(), i = 0;
+        while (i < len) {
+            while (i < len && view.byteAt(i) == ' ') i++;
+            int start = i;
+            while (i < len && view.byteAt(i) != ',') i++;
+            if (tokenEqualsIgnoreCase(view, start, i, token)) return true;
+            i++;
+        }
+        return false;
+    }
+
+    /** Case-insensitive compare of {@code view[start, end)}, trimming trailing spaces, against {@code token}. */
+    public static boolean tokenEqualsIgnoreCase(ByteView view, int start, int end, String token) {
+        int tlen = token.length();
+        int wlen = end - start;
+        while (wlen > 0 && view.byteAt(start + wlen - 1) == ' ') wlen--;
+        if (wlen != tlen) return false;
+        for (int i = 0; i < tlen; i++) {
+            byte b = view.byteAt(start + i);
+            if (b >= 'A' && b <= 'Z') b += 32;
+            if (b != (byte) token.charAt(i)) return false;
+        }
+        return true;
+    }
+}

flash/src/main/java/dev/relism/flash/http1/Http1ResponseWriter.java

@@ -0,0 +1,238 @@
+package dev.relism.flash.http1;
+
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http.DateHeader;
+import dev.relism.flash.http.Http1Limits;
+import dev.relism.flash.http.HttpMethod;
+import dev.relism.flash.http.HttpStatus;
+import dev.relism.flash.models.Response;
+import dev.relism.flash.transport.ConnectionScratch;
+
+import java.io.IOException;
+import java.io.InputStream;
+import java.io.OutputStream;
+import java.nio.charset.StandardCharsets;
+
+/**
+ * Serializes a {@link Response} as an HTTP/1.1 message. Sole responsibility: response
+ * serialization — routing, handler dispatch, and the request loop live in
+ * {@link Http1Connection}.
+ *
+ * The status line, {@code Content-Type}, {@code Date}, every custom header, and
+ * {@code Content-Length}/{@code Connection} are all serialized into
+ * {@link ConnectionScratch#responseHead} (a reused {@link ByteWriter}) before a single
+ * {@code OutputStream.write} call — not one small {@code write} per field, and no
+ * {@link java.io.BufferedOutputStream} coalescing them at the stream layer (this class removes
+ * the need for one entirely on the h1 response path). A body at or below
+ * {@link Http1Limits#INLINE_BODY_THRESHOLD} is copied into the same scratch buffer and goes out
+ * in that same syscall; a larger body is written separately right after, since copying it into
+ * the head buffer first would cost an extra full-body memcpy the syscall it saves does not pay
+ * for. Streaming/chunked bodies write the head, then relay their own bytes as they arrive — by
+ * definition unknown or too large to fold into one buffer up front.
+ */
+public final class Http1ResponseWriter {
+
+    private Http1ResponseWriter() {
+    }
+
+    private static final byte[] HTTP_1_1              = "HTTP/1.1 ".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] CRLF                  = "\r\n".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] CONTENT_TYPE          = "Content-Type: ".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] CONTENT_LENGTH        = "Content-Length: ".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] CONNECTION_CLOSE      = "Connection: close\r\n".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] CONNECTION_KEEPALIVE  = "Connection: keep-alive\r\n".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] TRANSFER_CHUNKED      = "Transfer-Encoding: chunked\r\n".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] FINAL_CHUNK           = "0\r\n\r\n".getBytes(StandardCharsets.UTF_8);
+    private static final byte[] UNKNOWN_STATUS_SUFFIX = " Unknown".getBytes(StandardCharsets.UTF_8);
+
+    /**
+     * Writes {@code response} to {@code out} as a complete HTTP/1.1 message.
+     *
+     * @param method   the request method — {@code null} is treated as "not HEAD" (used for
+     *                  parser-rejection responses, which never reach a handler and so have no
+     *                  associated method)
+     * @param sendDate whether to include the {@code Date} header ({@code FlashConfiguration#isSendDate()})
+     */
+    public static void writeResponse(OutputStream out, Response response, HttpMethod method,
+                                      boolean keepAlive, boolean sendDate, ConnectionScratch scratch) throws IOException {
+        int statusCode = response.getStatusCode();
+        // RFC 9110 §8.6/§15: 204, 304 and all 1xx responses MUST NOT carry Content-Length or a
+        // still reports the Content-Length GET would have, but never writes body bytes.
+        boolean noContentAllowed = statusCode == 204 || statusCode == 304 || (statusCode >= 100 && statusCode < 200);
+        boolean suppressBody     = noContentAllowed || method == HttpMethod.HEAD;
+
+        ByteWriter head = scratch.responseHead;
+        head.reset();
+        head.writeBytes(HTTP_1_1);
+        byte[] statusBytes = response.getStatusBytes();
+        if (statusBytes != null) head.writeBytes(statusBytes);
+        else writeStatusPhrase(head, statusCode);
+        head.writeBytes(CRLF);
+
+        // "Content-Type: \r\n" — a header with no value. Skip the line entirely instead.
+        byte[] contentType = response.getContentType();
+        if (contentType != null && contentType.length > 0) {
+            head.writeBytes(CONTENT_TYPE);
+            head.writeBytes(contentType);
+            head.writeBytes(CRLF);
+        }
+
+        // one write into the scratch, never a per-response format call.
+        if (sendDate) head.writeBytes(DateHeader.bytes());
+
+        response.writeHeadersInto(head);
+
+        if (response.hasTrailers()) {
+            writeTrailerBody(out, head, response, keepAlive, suppressBody, scratch);
+        } else if (response.isStreaming()) {
+            writeStreamingBody(out, head, response, keepAlive, noContentAllowed, suppressBody, scratch);
+        } else {
+            byte[] body = response.getBody();
+            int len = body != null ? body.length : 0;
+            if (!noContentAllowed) {
+                head.writeBytes(CONTENT_LENGTH);
+                head.writeDecimal(len);
+                head.writeBytes(CRLF);
+            }
+            head.writeBytes(keepAlive ? CONNECTION_KEEPALIVE : CONNECTION_CLOSE);
+            head.writeBytes(CRLF);
+
+            // the body itself.
+            boolean writeBody = body != null && !suppressBody;
+            if (writeBody && len <= Http1Limits.INLINE_BODY_THRESHOLD) {
+                // one syscall.
+                head.writeBytes(body);
+                out.write(head.array(), 0, head.length());
+            } else {
+                out.write(head.array(), 0, head.length());
+                if (writeBody) out.write(body);
+            }
+        }
+        out.flush();
+    }
+
+    private static void writeTrailerBody(OutputStream out, ByteWriter head, Response response,
+                                         boolean keepAlive, boolean suppressBody,
+                                         ConnectionScratch scratch) throws IOException {
+        head.writeBytes(TRANSFER_CHUNKED);
+        head.writeBytes(keepAlive ? CONNECTION_KEEPALIVE : CONNECTION_CLOSE);
+        head.writeBytes(CRLF);
+        out.write(head.array(), 0, head.length());
+        if (suppressBody) return;
+        if (response.isStreaming()) {
+            writeChunkedAndClose(out, response, scratch);
+        } else {
+            byte[] body = response.getBody();
+            if (body != null && body.length != 0) {
+                writeHex(out, body.length);
+                out.write(CRLF);
+                out.write(body);
+                out.write(CRLF);
+            }
+            writeFinalChunk(out, response);
+        }
+    }
+
+    private static void writeStreamingBody(OutputStream out, ByteWriter head, Response response, boolean keepAlive,
+                                            boolean noContentAllowed, boolean suppressBody,
+                                            ConnectionScratch scratch) throws IOException {
+        if (!response.isChunked()) {
+            if (!noContentAllowed) {
+                head.writeBytes(CONTENT_LENGTH);
+                head.writeDecimal(response.getStreamLength());
+                head.writeBytes(CRLF);
+            }
+            head.writeBytes(keepAlive ? CONNECTION_KEEPALIVE : CONNECTION_CLOSE);
+            head.writeBytes(CRLF);
+            out.write(head.array(), 0, head.length());
+            if (!suppressBody) relayAndClose(response.getStream(), out, scratch);
+        } else {
+            head.writeBytes(TRANSFER_CHUNKED);
+            head.writeBytes(keepAlive ? CONNECTION_KEEPALIVE : CONNECTION_CLOSE);
+            head.writeBytes(CRLF);
+            out.write(head.array(), 0, head.length());
+            // A HEAD response still declares the Transfer-Encoding GET would have used (RFC
+            // 9110 §9.3.2) but writes zero body bytes — not even the final-chunk marker, since
+            // there is no chunk framing at all for a message with no body.
+            if (!suppressBody) writeChunkedAndClose(out, response, scratch);
+        }
+    }
+
+    /**
+     * Closes the handler's stream on every exit — clean EOF or a write failure partway through
+     * (e.g. the client disconnected mid-transfer). Without this, a handler whose stream only
+     * releases a held resource (a pooled backend connection, say) from {@code close()} — not from
+     * observing EOF on a {@code read()} that a downstream write failure means it never reaches —
+     * leaks that resource for as long as the JVM takes to finalize it. A well-behaved stream's
+     * {@code close()} must already be idempotent (Java's own contract for {@link InputStream}), so
+     * this costs nothing extra on the ordinary clean-EOF path.
+     */
+    private static void relayAndClose(InputStream in, OutputStream out, ConnectionScratch scratch)
+            throws IOException {
+        try {
+            relay(in, out, scratch);
+        } finally {
+            in.close();
+        }
+    }
+
+    private static void writeChunkedAndClose(OutputStream out, Response response, ConnectionScratch scratch)
+            throws IOException {
+        try {
+            writeChunked(out, response.getStream(), response, scratch);
+        } finally {
+            response.getStream().close();
+        }
+    }
+
+    /**
+     * Copies {@code in} to {@code out} until EOF, via {@link ConnectionScratch#relayBuffer}
+     * instead of a fresh {@code byte[]} per call.
+     */
+    private static void relay(InputStream in, OutputStream out, ConnectionScratch scratch) throws IOException {
+        byte[] buf = scratch.relayBuffer;
+        int n;
+        while ((n = in.read(buf)) > 0) out.write(buf, 0, n);
+    }
+
+    private static void writeStatusPhrase(ByteWriter head, int statusCode) {
+        byte[] phrase = HttpStatus.bytesForCode(statusCode);
+        if (phrase != null) head.writeBytes(phrase);
+        else { head.writeDecimal(statusCode); head.writeBytes(UNKNOWN_STATUS_SUFFIX); }
+    }
+
+    private static void writeChunked(OutputStream out, InputStream stream, Response response,
+                                     ConnectionScratch scratch) throws IOException {
+        byte[] buf = scratch.relayBuffer;
+        int n;
+        while ((n = stream.read(buf)) > 0) {
+            writeHex(out, n);
+            out.write(CRLF);
+            out.write(buf, 0, n);
+            out.write(CRLF);
+        }
+        if (response.hasTrailers()) writeFinalChunk(out, response);
+        else out.write(FINAL_CHUNK);
+    }
+
+    private static void writeFinalChunk(OutputStream out, Response response) throws IOException {
+        out.write('0');
+        out.write(CRLF);
+        response.writeTrailers(out);
+        out.write(CRLF);
+    }
+
+    private static void writeHex(OutputStream out, int value) throws IOException {
+        int shift = 28;
+        boolean leading = true;
+        while (shift >= 0) {
+            int digit = (value >>> shift) & 0xF;
+            if (digit != 0 || !leading) {
+                leading = false;
+                out.write(digit < 10 ? '0' + digit : 'a' + digit - 10);
+            }
+            shift -= 4;
+        }
+        if (leading) out.write('0');
+    }
+}

flash/src/main/java/dev/relism/flash/http2/Http2AbuseGuard.java

@@ -0,0 +1,114 @@
+package dev.relism.flash.http2;
+
+import dev.relism.flash.extension.FlashConfiguration;
+
+/** Enforces per-connection HTTP/2 rate limits and lifetime budgets. */
+final class Http2AbuseGuard {
+  private RollingWindowCounter resetRate;
+  private RollingWindowCounter streamCreationRate;
+  private RollingWindowCounter settingsRate;
+  private RollingWindowCounter pingRate;
+  private RollingWindowCounter uselessFrameRate;
+  private int maxResetRate;
+  private int maxStreamCreationRate;
+  private long maxStreams;
+  private long maxBytes;
+  private long maxLifetimeNanos;
+  private long startedNanos;
+  private long wireBytes;
+  private long streams;
+
+  Http2AbuseGuard() {
+    configure(FlashConfiguration.builder().build());
+  }
+
+  void configure(FlashConfiguration configuration) {
+    long interval = configuration.getH2AbuseRateIntervalMs();
+    if (interval < 2) throw new IllegalArgumentException("h2AbuseRateIntervalMs must be at least 2");
+    resetRate = new RollingWindowCounter(interval);
+    streamCreationRate = new RollingWindowCounter(interval);
+    settingsRate = new RollingWindowCounter(interval);
+    pingRate = new RollingWindowCounter(interval);
+    uselessFrameRate = new RollingWindowCounter(interval);
+    maxResetRate =
+        positive(
+            configuration.getH2MaxResetStreamsPerInterval(),
+            "h2MaxResetStreamsPerInterval");
+    maxStreamCreationRate =
+        positive(
+            configuration.getH2MaxStreamsCreatedPerInterval(),
+            "h2MaxStreamsCreatedPerInterval");
+    maxStreams =
+        nonNegative(configuration.getH2MaxStreamsPerConnection(), "h2MaxStreamsPerConnection");
+    maxBytes =
+        nonNegative(configuration.getH2MaxBytesPerConnection(), "h2MaxBytesPerConnection");
+    long lifetime =
+        nonNegative(
+            configuration.getH2MaxConnectionLifetimeMs(), "h2MaxConnectionLifetimeMs");
+    maxLifetimeNanos = toNanos(lifetime);
+  }
+
+  void start() {
+    startedNanos = System.nanoTime();
+  }
+
+  void receivedFrame(int payloadLength) {
+    wireBytes += 9L + payloadLength;
+    checkBudgets();
+  }
+
+  void streamCreated() {
+    if (streamCreationRate.incrementExceeded(maxStreamCreationRate)) {
+      calm("stream creation rate");
+    }
+    streams++;
+    if (maxStreams > 0 && streams > maxStreams) calm("connection stream budget");
+  }
+
+  void resetReceived() {
+    if (resetRate.incrementExceeded(maxResetRate)) calm("RST_STREAM rate");
+  }
+
+  void settingsReceived() {
+    if (settingsRate.incrementExceeded(Http2Limits.MAX_SETTINGS_PER_INTERVAL)) {
+      calm("SETTINGS rate");
+    }
+  }
+
+  void pingReceived() {
+    if (pingRate.incrementExceeded(Http2Limits.MAX_PINGS_PER_INTERVAL)) calm("PING rate");
+  }
+
+  void uselessFrameReceived() {
+    if (uselessFrameRate.incrementExceeded(Http2Limits.MAX_USELESS_FRAMES_PER_INTERVAL)) {
+      calm("non-progress frame rate");
+    }
+  }
+
+  void checkBudgets() {
+    if (maxBytes > 0 && wireBytes > maxBytes) calm("connection byte budget");
+    if (maxLifetimeNanos > 0 && System.nanoTime() - startedNanos > maxLifetimeNanos) {
+      calm("connection lifetime budget");
+    }
+  }
+
+  private static int positive(int value, String name) {
+    if (value <= 0) throw new IllegalArgumentException(name + " must be positive");
+    return value;
+  }
+
+  private static long nonNegative(long value, String name) {
+    if (value < 0) throw new IllegalArgumentException(name + " must not be negative");
+    return value;
+  }
+
+  private static long toNanos(long milliseconds) {
+    return milliseconds > Long.MAX_VALUE / 1_000_000L
+        ? Long.MAX_VALUE
+        : milliseconds * 1_000_000L;
+  }
+
+  private static void calm(String reason) {
+    throw Http2Exception.of(Http2ErrorCode.ENHANCE_YOUR_CALM, reason + " exceeded");
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2Authority.java

@@ -0,0 +1,54 @@
+package dev.relism.flash.http2;
+
+import java.security.cert.Certificate;
+import java.security.cert.CertificateParsingException;
+import java.security.cert.X509Certificate;
+import java.util.Collection;
+import java.util.List;
+import java.util.Locale;
+import javax.net.ssl.SSLSession;
+
+/** Validates a coalesced request authority against the certificate selected for its connection. */
+final class Http2Authority {
+  private Http2Authority() {}
+
+  static boolean isServed(String authority, SSLSession session) {
+    if (session == null || authority == null) return true;
+    String host = host(authority);
+    try {
+      Certificate[] certificates = session.getLocalCertificates();
+      if (certificates == null || certificates.length == 0
+          || !(certificates[0] instanceof X509Certificate certificate)) {
+        return true;
+      }
+      Collection<List<?>> names = certificate.getSubjectAlternativeNames();
+      if (names == null) return true;
+      for (List<?> name : names) {
+        int type = (Integer) name.get(0);
+        if ((type == 2 || type == 7) && matches(host, name.get(1).toString())) return true;
+      }
+      return false;
+    } catch (CertificateParsingException failure) {
+      return true;
+    }
+  }
+
+  static boolean matches(String authority, String certificateName) {
+    String host = host(authority).toLowerCase(Locale.ROOT);
+    String name = certificateName.toLowerCase(Locale.ROOT);
+    if (!name.startsWith("*.")) return host.equals(name);
+    String suffix = name.substring(1);
+    if (!host.endsWith(suffix)) return false;
+    int prefixLength = host.length() - suffix.length();
+    return prefixLength > 0 && host.indexOf('.') == prefixLength;
+  }
+
+  private static String host(String authority) {
+    if (authority.startsWith("[")) {
+      int closing = authority.indexOf(']');
+      return closing < 0 ? authority : authority.substring(1, closing);
+    }
+    int colon = authority.lastIndexOf(':');
+    return colon > 0 && authority.indexOf(':') == colon ? authority.substring(0, colon) : authority;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2Connection.java

@@ -0,0 +1,660 @@
+package dev.relism.flash.http2;
+
+import dev.relism.flash.bytes.Pairs;
+import dev.relism.flash.extension.FlashConfiguration;
+import dev.relism.flash.http2.Http2ConnectionScratch.ControlIntent;
+import dev.relism.flash.http2.Http2ConnectionScratch.ControlKind;
+import dev.relism.flash.http2.frame.FrameFlags;
+import dev.relism.flash.http2.frame.FrameHeader;
+import dev.relism.flash.http2.frame.FrameType;
+import dev.relism.flash.http2.frame.FrameValidator;
+import dev.relism.flash.http2.frame.Http2FrameReader;
+import dev.relism.flash.http2.frame.Http2FrameWriter;
+import dev.relism.flash.http2.frame.Padding;
+import dev.relism.flash.http2.hpack.HeaderSink;
+import dev.relism.flash.http2.message.DataBufferPool;
+import dev.relism.flash.http2.stream.Http2FlowController;
+import dev.relism.flash.http2.stream.Http2Stream;
+import dev.relism.flash.http2.stream.Http2StreamState;
+import dev.relism.flash.http2.stream.Http2StreamTable;
+import dev.relism.flash.transport.BufferedByteSource;
+import dev.relism.flash.transport.ConnectionContext;
+import dev.relism.flash.transport.ConnectionProtocol;
+import java.io.IOException;
+import java.net.SocketTimeoutException;
+import java.util.function.BooleanSupplier;
+import lombok.extern.slf4j.Slf4j;
+
+/**
+ * Owns one HTTP/2 connection's demultiplexing and connection-level protocol state. The demux loop
+ * never invokes application code and never waits for a handler or body consumer; stream dispatch is
+ * handed to independent virtual threads by the stream layer.
+ */
+@Slf4j
+public final class Http2Connection implements ConnectionProtocol {
+  private static final byte[] SHUTDOWN_PING = {
+    (byte) 0x46, (byte) 0x4c, (byte) 0x41, (byte) 0x53,
+    (byte) 0x48, (byte) 0x47, (byte) 0x4f, (byte) 0x21
+  };
+
+  private final Http2Settings peerSettings = new Http2Settings();
+  private final Http2ConnectionScratch scratch = new Http2ConnectionScratch();
+  private final Http2Settings.StreamWindowUpdater streamWindows;
+  private final long settingsAckTimeoutMs;
+  private final Http2HeaderBlockDecoder headerBlocks = new Http2HeaderBlockDecoder();
+  private final DataBufferPool dataBuffers =
+      new DataBufferPool(Http2Limits.MAX_FRAME_SIZE_LOCAL, Http2Limits.DATA_BUFFER_POOL_SIZE);
+  private final Http2StreamTable streams =
+      new Http2StreamTable(Http2Limits.MAX_CONCURRENT_STREAMS, dataBuffers);
+  private static final HeaderSink DISCARD_HEADERS = (name, value, never) -> {};
+
+  private Http2FlowController flowController;
+  private int outstandingLocalSettings;
+  private long oldestSettingsSentNanos;
+  private int lastProcessedStreamId;
+  private int peerLastStreamId = Integer.MAX_VALUE;
+  private int peerErrorCode;
+  private boolean peerGoAway;
+  private boolean gracefulStarted;
+  private boolean gracefulFinished;
+  private int highestClientStreamId;
+  private Http2Stream pendingHeaderStream;
+  private boolean refusingHeaderStream;
+  private boolean pendingTrailers;
+  private Http2StreamDispatcher streamDispatcher;
+  private final Http2Stream[] dispatchQueue = new Http2Stream[Http2Limits.MAX_CONCURRENT_STREAMS];
+  private int dispatchCount;
+  private final Http2AbuseGuard abuse = new Http2AbuseGuard();
+  private long streamIdleTimeoutNanos = Http2Limits.STREAM_IDLE_TIMEOUT_MS * 1_000_000L;
+  private final Http2Stream[] idleSweep = new Http2Stream[Http2Limits.MAX_CONCURRENT_STREAMS];
+
+  public Http2Connection() {
+    this(delta -> {}, Http2Limits.SETTINGS_ACK_TIMEOUT_MS);
+  }
+
+  public Http2Connection(Http2Settings.StreamWindowUpdater streamWindows) {
+    this(streamWindows, Http2Limits.SETTINGS_ACK_TIMEOUT_MS);
+  }
+
+  Http2Connection(Http2Settings.StreamWindowUpdater streamWindows, long settingsAckTimeoutMs) {
+    this.streamWindows =
+        delta -> {
+          try {
+            if (flowController == null) streams.adjustAllSendWindows(delta);
+            else flowController.applyInitialWindowDelta(streams, delta);
+          } catch (IllegalStateException overflow) {
+            throw Http2Exception.FLOW_CONTROL_ERROR;
+          }
+          streamWindows.applyInitialWindowDelta(delta);
+        };
+    this.settingsAckTimeoutMs = settingsAckTimeoutMs;
+  }
+
+  @Override
+  public void run(ConnectionContext ctx) throws IOException {
+    configure(ctx.configuration());
+    Http2FrameWriter writer = new Http2FrameWriter(ctx.rawOut()::write);
+    flowController =
+        new Http2FlowController(
+            (streamId, increment) -> sendWindowUpdate(writer, streamId, increment));
+    streamDispatcher =
+        new Http2StreamDispatcher(
+            ctx,
+            writer,
+            peerSettings,
+            streams,
+            flowController,
+            (streamId, error) -> sendRstStream(writer, streamId, error));
+    try {
+      run(ctx.in(), writer, ctx.stopped());
+    } finally {
+      writer.close();
+    }
+  }
+
+  void run(BufferedByteSource input, Http2FrameWriter writer, BooleanSupplier stopped)
+      throws IOException {
+    if (flowController == null) {
+      flowController =
+          new Http2FlowController(
+              (streamId, increment) -> sendWindowUpdate(writer, streamId, increment));
+    }
+    Http2FrameReader reader = new Http2FrameReader(input);
+    runPrepared(input, reader, writer, stopped);
+  }
+
+  /** Runs with connection collaborators that were allocated during connection setup. */
+  void runPrepared(
+      BufferedByteSource input,
+      Http2FrameReader reader,
+      Http2FrameWriter writer,
+      BooleanSupplier stopped)
+      throws IOException {
+    PrefaceResult preface = verifyPreface(input);
+    if (preface == PrefaceResult.TRUNCATED) return;
+    if (preface == PrefaceResult.INVALID) {
+      sendGoAway(writer, 0, Http2ErrorCode.PROTOCOL_ERROR, "invalid client preface");
+      writer.drain();
+      return;
+    }
+    abuse.start();
+
+    sendConstant(writer, Http2Preface.serverSettings());
+    sendConstant(writer, Http2Preface.initialConnectionWindow());
+    outstandingLocalSettings = 1;
+    oldestSettingsSentNanos = System.nanoTime();
+
+    boolean firstFrame = true;
+    try {
+      while (!gracefulFinished && !peerGoAway) {
+        abuse.checkBudgets();
+        closeIdleStreams(writer);
+        if (stopped.getAsBoolean() && !gracefulStarted) startGracefulShutdown(writer);
+        FrameHeader frame;
+        try {
+          frame = reader.readFrame(Math.min(100, nextReadTimeoutMs()));
+        } catch (SocketTimeoutException timeout) {
+          checkSettingsTimeout();
+          headerBlocks.checkTimeout();
+          if (stopped.getAsBoolean() && !gracefulStarted) {
+            startGracefulShutdown(writer);
+            continue;
+          }
+          if (reader.frameDeadlineExpired()) throw timeout;
+          continue;
+        }
+        if (frame == null) break;
+        abuse.receivedFrame(frame.length());
+        try {
+          headerBlocks.checkTimeout();
+          FrameValidator.validate(frame, headerBlocks.insideHeaderBlock());
+          if (headerBlocks.insideHeaderBlock() && frame.type() != FrameType.CONTINUATION) {
+            throw Http2Exception.PROTOCOL_ERROR;
+          }
+          if (firstFrame && frame.type() != FrameType.SETTINGS) {
+            throw Http2Exception.PROTOCOL_ERROR;
+          }
+          if (firstFrame && FrameFlags.isAck(frame.flags()) && frame.length() == 0) {
+            throw Http2Exception.PROTOCOL_ERROR;
+          }
+          firstFrame = false;
+          dispatch(frame, writer);
+        } catch (Http2StreamException streamError) {
+          sendRstStream(writer, streamError);
+          closeStreamAfterError(streamError.streamId());
+        } finally {
+          reader.consumeFrame();
+        }
+        writer.drain();
+        if (dispatchCount > 0 && !reader.hasBufferedInput()) dispatchPendingStreams();
+        checkSettingsTimeout();
+      }
+    } catch (Http2Exception connectionError) {
+      sendGoAway(
+          writer, lastProcessedStreamId, connectionError.errorCode(), connectionError.getMessage());
+    } catch (IOException io) {
+      throw io;
+    } catch (RuntimeException unexpected) {
+      log.error("Unexpected failure in HTTP/2 demux loop", unexpected);
+      sendGoAway(writer, lastProcessedStreamId, Http2ErrorCode.INTERNAL_ERROR, "internal error");
+    }
+  }
+
+  private PrefaceResult verifyPreface(BufferedByteSource input) throws IOException {
+    byte[] preface = scratch.prefaceBuffer();
+    int read = 0;
+    input.setDeadline(System.nanoTime() + Http2Limits.FRAME_READ_TIMEOUT_MS * 1_000_000L);
+    try {
+      while (read < preface.length) {
+        int n = input.read(preface, read, preface.length - read);
+        if (n < 0) return PrefaceResult.TRUNCATED;
+        read += n;
+      }
+      return Http2Preface.matchesClientPreface(preface)
+          ? PrefaceResult.MATCHED
+          : PrefaceResult.INVALID;
+    } finally {
+      input.clearDeadline();
+    }
+  }
+
+  private enum PrefaceResult {
+    MATCHED,
+    INVALID,
+    TRUNCATED
+  }
+
+  private void dispatch(FrameHeader frame, Http2FrameWriter writer) throws IOException {
+    FrameType type = frame.type();
+    if (type == null) {
+      abuse.uselessFrameReceived();
+      return;
+    }
+    switch (type) {
+      case SETTINGS -> receiveSettings(frame, writer);
+      case PING -> receivePing(frame, writer);
+      case WINDOW_UPDATE -> receiveWindowUpdate(frame);
+      case GOAWAY -> receiveGoAway(frame);
+      case HEADERS -> receiveHeaders(frame, writer);
+      case CONTINUATION -> receiveContinuation(frame, writer);
+      case DATA -> receiveData(frame);
+      case RST_STREAM -> receiveRstStream(frame);
+      case PRIORITY -> receivePriority(frame);
+      default -> {}
+    }
+  }
+
+  private void receiveHeaders(FrameHeader frame, Http2FrameWriter writer) throws IOException {
+    int streamId = frame.streamId();
+    if ((streamId & 1) == 0) throw Http2Exception.PROTOCOL_ERROR;
+    Http2Stream existing = streams.get(streamId);
+    if (existing != null) {
+      existing.touch();
+      if (existing.state() == Http2StreamState.HALF_CLOSED_REMOTE) {
+        throw new Http2StreamException(
+            streamId, Http2ErrorCode.STREAM_CLOSED, "stream is half-closed remotely");
+      }
+      if (!FrameFlags.isEndStream(frame.flags())) {
+        throw new Http2StreamException(
+            streamId, Http2ErrorCode.PROTOCOL_ERROR, "trailers require END_STREAM");
+      }
+      pendingHeaderStream = existing;
+      pendingTrailers = true;
+      existing.trailerBlock().reset();
+      if (headerBlocks.accept(frame, existing.trailerBlock())) completeHeaders(writer, streamId);
+      return;
+    }
+    if (streamId <= highestClientStreamId) {
+      int closedKind = streams.closedKind(streamId);
+      if (closedKind == Http2StreamTable.CLOSED_NORMALLY) {
+        throw Http2Exception.of(Http2ErrorCode.STREAM_CLOSED, "frame on a closed stream");
+      }
+      if (closedKind == Http2StreamTable.CLOSED_BY_RESET) {
+        throw new Http2StreamException(
+            streamId, Http2ErrorCode.STREAM_CLOSED, "stream was reset");
+      }
+      throw Http2Exception.PROTOCOL_ERROR;
+    }
+    abuse.streamCreated();
+    highestClientStreamId = streamId;
+    pendingTrailers = false;
+
+    pendingHeaderStream = streams.acquire(streamId);
+    refusingHeaderStream = pendingHeaderStream == null;
+    if (pendingHeaderStream != null) {
+      flowController.initializeStreamSendWindow(
+          pendingHeaderStream, peerSettings.initialWindowSize());
+    }
+    HeaderSink sink =
+        refusingHeaderStream
+            ? DISCARD_HEADERS
+            : (pendingTrailers
+                ? pendingHeaderStream.trailerBlock()
+                : pendingHeaderStream.headerBlock());
+    if (headerBlocks.accept(frame, sink)) completeHeaders(writer, frame.streamId());
+  }
+
+  private void receiveContinuation(FrameHeader frame, Http2FrameWriter writer) throws IOException {
+    if (pendingHeaderStream == null && !refusingHeaderStream) {
+      throw Http2Exception.PROTOCOL_ERROR;
+    }
+    HeaderSink sink =
+        refusingHeaderStream
+            ? DISCARD_HEADERS
+            : (pendingTrailers
+                ? pendingHeaderStream.trailerBlock()
+                : pendingHeaderStream.headerBlock());
+    if (headerBlocks.accept(frame, sink)) completeHeaders(writer, frame.streamId());
+  }
+
+  private void completeHeaders(Http2FrameWriter writer, int streamId) throws IOException {
+    try {
+      if (refusingHeaderStream) {
+        sendRstStream(writer, streamId, Http2ErrorCode.REFUSED_STREAM);
+      } else {
+        Http2Stream stream = pendingHeaderStream;
+        boolean dispatch;
+        if (pendingTrailers) {
+          stream.validateTrailers();
+          stream.finishRequestBody();
+          stream.transition(Http2StreamState.Event.RECV_HEADERS_ES);
+          dispatch = !stream.dispatched();
+        } else {
+          if (streamDispatcher != null) stream.validateHeaders();
+          dispatch = stream.prepareRequestBody(flowController, headerBlocks.endStream());
+          stream.transition(
+              headerBlocks.endStream()
+                  ? Http2StreamState.Event.RECV_HEADERS_ES
+                  : Http2StreamState.Event.RECV_HEADERS);
+          lastProcessedStreamId = Math.max(lastProcessedStreamId, streamId);
+        }
+        if (streamDispatcher == null && !pendingTrailers) {
+          streams.retire(stream, streamId);
+          if (!gracefulStarted) startGracefulShutdown(writer);
+        } else if (dispatch) {
+          enqueueDispatch(stream);
+        } else if (stream.responseStarted() && stream.responseWriter().finished()
+            && !stream.responseInFlight() && stream.state() == Http2StreamState.CLOSED) {
+          streams.retire(stream, streamId);
+        }
+      }
+    } finally {
+      pendingHeaderStream = null;
+      refusingHeaderStream = false;
+      pendingTrailers = false;
+    }
+  }
+
+  private void receivePriority(FrameHeader frame) {
+    abuse.uselessFrameReceived();
+    int dependency = readUInt31(frame.buffer(), frame.payloadOffset());
+    if (dependency == frame.streamId()) {
+      throw new Http2StreamException(
+          frame.streamId(), Http2ErrorCode.PROTOCOL_ERROR, "stream depends on itself");
+    }
+  }
+
+  private void receiveData(FrameHeader frame) {
+    if (frame.length() == 0) abuse.uselessFrameReceived();
+    flowController.receiveConnectionBytes(frame.length());
+    Http2Stream stream = streams.get(frame.streamId());
+    if (stream == null) {
+      discardConnectionBytes(frame.length());
+      if (frame.streamId() > highestClientStreamId) throw Http2Exception.PROTOCOL_ERROR;
+      throw new Http2StreamException(
+          frame.streamId(), Http2ErrorCode.STREAM_CLOSED, "stream is closed");
+    }
+    boolean bodyAccepted = false;
+    try {
+      stream.touch();
+      stream.transition(
+          FrameFlags.isEndStream(frame.flags())
+              ? Http2StreamState.Event.RECV_DATA_ES
+              : Http2StreamState.Event.RECV_DATA);
+      flowController.receiveStreamBytes(stream, frame.length());
+      long unpadded =
+          Padding.unpad(
+              frame.buffer(),
+              frame.payloadOffset(),
+              frame.length(),
+              FrameFlags.isPadded(frame.flags()));
+      int dataOffset = Pairs.hi(unpadded);
+      int dataLength = Pairs.lo(unpadded);
+      if (frame.length() == 0) {
+        if (stream.incrementEmptyDataFrames()
+            > Http2Limits.MAX_EMPTY_DATA_FRAMES_PER_STREAM) {
+          throw new Http2StreamException(
+              frame.streamId(), Http2ErrorCode.ENHANCE_YOUR_CALM, "empty DATA frame limit exceeded");
+        }
+      } else {
+        stream.resetEmptyDataFrames();
+      }
+      stream.receiveData(frame.buffer(), dataOffset, dataLength, frame.length());
+      bodyAccepted = true;
+      if (FrameFlags.isEndStream(frame.flags())) {
+        stream.finishRequestBody();
+        if (streamDispatcher != null && !stream.dispatched()) enqueueDispatch(stream);
+        else if (stream.responseStarted() && stream.responseWriter().finished()
+            && !stream.responseInFlight()
+            && stream.state() == Http2StreamState.CLOSED) {
+          streams.retire(stream, frame.streamId());
+        }
+      }
+    } catch (RuntimeException failure) {
+      if (!bodyAccepted) discardConnectionBytes(frame.length());
+      throw failure;
+    }
+  }
+
+  private void receiveRstStream(FrameHeader frame) {
+    abuse.resetReceived();
+    Http2Stream stream = streams.get(frame.streamId());
+    if (stream == null) {
+      if (frame.streamId() > highestClientStreamId) throw Http2Exception.PROTOCOL_ERROR;
+      return;
+    }
+    boolean releaseDeferred =
+        stream.dispatched() || stream.state() == Http2StreamState.HALF_CLOSED_REMOTE;
+    stream.transition(Http2StreamState.Event.RECV_RST);
+    if (!streams.removeIfSame(stream, frame.streamId())) return;
+    streams.rememberReset(frame.streamId());
+    if (releaseDeferred) {
+      stream.cancel();
+      if (stream.responseStarted() && !stream.responseInFlight()) streams.release(stream);
+    } else {
+      streams.release(stream);
+    }
+  }
+
+  private void enqueueDispatch(Http2Stream stream) {
+    if (dispatchCount == dispatchQueue.length) {
+      throw new Http2StreamException(
+          stream.id(), Http2ErrorCode.REFUSED_STREAM, "dispatch queue is full");
+    }
+    stream.markDispatched();
+    dispatchQueue[dispatchCount++] = stream;
+  }
+
+  private void dispatchPendingStreams() {
+    int count = dispatchCount;
+    dispatchCount = 0;
+    for (int i = 0; i < count; i++) {
+      Http2Stream stream = dispatchQueue[i];
+      dispatchQueue[i] = null;
+      streamDispatcher.dispatch(stream);
+    }
+  }
+
+  private void receiveSettings(FrameHeader frame, Http2FrameWriter writer) throws IOException {
+    boolean ack = FrameFlags.isAck(frame.flags());
+    if (ack) {
+      if (frame.length() != 0) throw Http2Exception.FRAME_SIZE_ERROR;
+      if (outstandingLocalSettings == 0) throw Http2Exception.PROTOCOL_ERROR;
+      outstandingLocalSettings--;
+      if (outstandingLocalSettings == 0) oldestSettingsSentNanos = 0;
+      return;
+    }
+    abuse.settingsReceived();
+    peerSettings.apply(frame.buffer(), frame.payloadOffset(), frame.length(), streamWindows);
+    sendConstant(writer, Http2Preface.settingsAck());
+  }
+
+  private void receivePing(FrameHeader frame, Http2FrameWriter writer) throws IOException {
+    if (FrameFlags.isAck(frame.flags())) {
+      if (gracefulStarted && matches(frame.buffer(), frame.payloadOffset(), SHUTDOWN_PING)) {
+        sendGoAway(writer, lastProcessedStreamId, Http2ErrorCode.NO_ERROR, "shutdown complete");
+        gracefulFinished = true;
+      }
+      return;
+    }
+    abuse.pingReceived();
+    ControlIntent pong = scratch.acquire(ControlKind.PING);
+    pong.frame(FrameType.PING, FrameFlags.ACK, 0, frame.buffer(), frame.payloadOffset(), 8);
+    writer.writePriority(pong);
+  }
+
+  private void receiveWindowUpdate(FrameHeader frame) {
+    abuse.uselessFrameReceived();
+    int increment = readUInt31(frame.buffer(), frame.payloadOffset());
+    if (increment == 0) {
+      if (frame.streamId() == 0) throw Http2Exception.PROTOCOL_ERROR;
+      throw new Http2StreamException(
+          frame.streamId(), Http2ErrorCode.PROTOCOL_ERROR, "zero window increment");
+    }
+    if (frame.streamId() != 0) {
+      Http2Stream stream = streams.get(frame.streamId());
+      if (stream == null) {
+        if (frame.streamId() > highestClientStreamId) throw Http2Exception.PROTOCOL_ERROR;
+        return;
+      }
+      try {
+        stream.touch();
+        flowController.increaseStreamSendWindow(stream, increment);
+      } catch (IllegalStateException overflow) {
+        throw new Http2StreamException(
+            frame.streamId(), Http2ErrorCode.FLOW_CONTROL_ERROR, "stream window overflow");
+      }
+      if (streamDispatcher != null) streamDispatcher.streamWindowUpdated(stream);
+      return;
+    }
+    flowController.increaseConnectionSendWindow(increment);
+    if (streamDispatcher != null) streamDispatcher.connectionWindowUpdated();
+  }
+
+  private void receiveGoAway(FrameHeader frame) {
+    peerLastStreamId = readUInt31(frame.buffer(), frame.payloadOffset());
+    peerErrorCode = readInt(frame.buffer(), frame.payloadOffset() + 4);
+    peerGoAway = true;
+  }
+
+  void configure(FlashConfiguration configuration) {
+    abuse.configure(configuration);
+    long idle = configuration.getH2StreamIdleTimeoutMs();
+    if (idle <= 0) throw new IllegalArgumentException("h2StreamIdleTimeoutMs must be positive");
+    streamIdleTimeoutNanos = idle > Long.MAX_VALUE / 1_000_000L
+        ? Long.MAX_VALUE : idle * 1_000_000L;
+  }
+
+  private void closeIdleStreams(Http2FrameWriter writer) throws IOException {
+    int count = streams.copyValues(idleSweep);
+    long now = System.nanoTime();
+    for (int i = 0; i < count; i++) {
+      Http2Stream stream = idleSweep[i];
+      idleSweep[i] = null;
+      if (!stream.idleExpired(now, streamIdleTimeoutNanos)) continue;
+      int streamId = stream.id();
+      if (!streams.removeIfSame(stream, streamId)) continue;
+      streams.rememberReset(streamId);
+      sendRstStream(writer, streamId, Http2ErrorCode.CANCEL);
+      if (stream.dispatched()) stream.cancel();
+      else streams.release(stream);
+    }
+  }
+
+  private void startGracefulShutdown(Http2FrameWriter writer) throws IOException {
+    gracefulStarted = true;
+    sendGoAway(writer, Integer.MAX_VALUE, Http2ErrorCode.NO_ERROR, "server shutting down");
+    ControlIntent ping = scratch.acquire(ControlKind.PING);
+    ping.frame(FrameType.PING, 0, 0, SHUTDOWN_PING, 0, SHUTDOWN_PING.length);
+    writer.writePriority(ping);
+  }
+
+  private void sendRstStream(Http2FrameWriter writer, Http2StreamException error)
+      throws IOException {
+    ControlIntent rst = scratch.acquire(ControlKind.SETTINGS_OR_OTHER);
+    rst.frame(FrameType.RST_STREAM, 0, error.streamId(), error.errorCode().bytes(), 0, 4);
+    writer.writePriority(rst);
+  }
+
+  private void sendRstStream(Http2FrameWriter writer, int streamId, Http2ErrorCode error)
+      throws IOException {
+    ControlIntent rst = scratch.acquire(ControlKind.SETTINGS_OR_OTHER);
+    rst.frame(FrameType.RST_STREAM, 0, streamId, error.bytes(), 0, 4);
+    writer.writePriority(rst);
+  }
+
+  private void sendWindowUpdate(Http2FrameWriter writer, int streamId, int increment)
+      throws IOException {
+    ControlIntent update = scratch.acquire(ControlKind.SETTINGS_OR_OTHER);
+    update.windowUpdate(streamId, increment);
+    writer.writePriority(update);
+  }
+
+  private void discardConnectionBytes(int bytes) {
+    try {
+      flowController.discarded(bytes);
+    } catch (IOException failure) {
+      throw new IllegalStateException("failed to restore connection flow-control window", failure);
+    }
+  }
+
+  private void closeStreamAfterError(int streamId) {
+    Http2Stream stream = streams.get(streamId);
+    if (stream == null) return;
+    if (!streams.removeIfSame(stream, streamId)) return;
+    streams.rememberReset(streamId);
+    if (stream.dispatched()) stream.cancel();
+    else streams.release(stream);
+    if (pendingHeaderStream == stream) pendingHeaderStream = null;
+  }
+
+  private void sendGoAway(
+      Http2FrameWriter writer, int lastStreamId, Http2ErrorCode error, String debug)
+      throws IOException {
+    ControlIntent goAway = scratch.acquire(ControlKind.SETTINGS_OR_OTHER);
+    goAway.goAway(lastStreamId, error, debug == null ? "" : debug);
+    writer.writePriority(goAway);
+  }
+
+  private void sendConstant(Http2FrameWriter writer, byte[] bytes) throws IOException {
+    ControlIntent intent = scratch.acquire(ControlKind.SETTINGS_OR_OTHER);
+    intent.copy(bytes);
+    writer.writePriority(intent);
+  }
+
+  private long nextReadTimeoutMs() {
+    if (outstandingLocalSettings == 0) return Http2Limits.FRAME_READ_TIMEOUT_MS;
+    long elapsed = System.nanoTime() - oldestSettingsSentNanos;
+    long remainingNanos = settingsAckTimeoutMs * 1_000_000L - elapsed;
+    if (remainingNanos <= 0) throw Http2Exception.SETTINGS_TIMEOUT;
+    long remainingMs = Math.max(1, (remainingNanos + 999_999L) / 1_000_000L);
+    return Math.min(Http2Limits.FRAME_READ_TIMEOUT_MS, remainingMs);
+  }
+
+  private void checkSettingsTimeout() {
+    if (outstandingLocalSettings != 0
+        && System.nanoTime() - oldestSettingsSentNanos >= settingsAckTimeoutMs * 1_000_000L) {
+      throw Http2Exception.SETTINGS_TIMEOUT;
+    }
+  }
+
+  private static boolean matches(byte[] buf, int off, byte[] expected) {
+    int different = 0;
+    for (int i = 0; i < expected.length; i++) different |= buf[off + i] ^ expected[i];
+    return different == 0;
+  }
+
+  private static int readUInt31(byte[] buf, int off) {
+    return readInt(buf, off) & 0x7FFF_FFFF;
+  }
+
+  private static int readInt(byte[] buf, int off) {
+    return ((buf[off] & 0xFF) << 24)
+        | ((buf[off + 1] & 0xFF) << 16)
+        | ((buf[off + 2] & 0xFF) << 8)
+        | (buf[off + 3] & 0xFF);
+  }
+
+  public Http2Settings peerSettings() {
+    return peerSettings;
+  }
+
+  public long connectionSendWindow() {
+    return flowController == null ? 65_535 : flowController.connectionSendWindow();
+  }
+
+  public int peerLastStreamId() {
+    return peerLastStreamId;
+  }
+
+  public int peerErrorCode() {
+    return peerErrorCode;
+  }
+
+  void reset() {
+    peerSettings.reset();
+    flowController = null;
+    outstandingLocalSettings = 0;
+    oldestSettingsSentNanos = 0;
+    lastProcessedStreamId = 0;
+    peerLastStreamId = Integer.MAX_VALUE;
+    peerErrorCode = 0;
+    peerGoAway = false;
+    gracefulStarted = false;
+    gracefulFinished = false;
+    highestClientStreamId = 0;
+    pendingHeaderStream = null;
+    refusingHeaderStream = false;
+    dispatchCount = 0;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2ConnectionScratch.java

@@ -0,0 +1,179 @@
+package dev.relism.flash.http2;
+
+import dev.relism.flash.http2.frame.FrameType;
+import dev.relism.flash.http2.frame.WriteIntent;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.atomic.AtomicInteger;
+
+/** Reusable control-frame storage owned by one HTTP/2 connection. */
+final class Http2ConnectionScratch {
+  private static final int CONTROL_SLOT_COUNT =
+      Http2Limits.MAX_PING_QUEUE_DEPTH + Http2Limits.MAX_SETTINGS_ACK_QUEUE_DEPTH + 8;
+  private static final int CONTROL_FRAME_CAPACITY = 256;
+
+  private final ControlIntent[] controls = new ControlIntent[CONTROL_SLOT_COUNT];
+  private final AtomicInteger pingResponses = new AtomicInteger();
+  private final AtomicInteger settingsAcks = new AtomicInteger();
+  private final byte[] preface = new byte[Http2Preface.clientPrefaceLength()];
+
+  Http2ConnectionScratch() {
+    for (int i = 0; i < controls.length; i++) {
+      controls[i] = new ControlIntent(this, CONTROL_FRAME_CAPACITY);
+    }
+  }
+
+  byte[] prefaceBuffer() {
+    return preface;
+  }
+
+  ControlIntent acquire(ControlKind kind) {
+    AtomicInteger counter = counter(kind);
+    int limit = limit(kind);
+    int queued = counter.incrementAndGet();
+    if (queued > limit) {
+      counter.decrementAndGet();
+      throw Http2Exception.of(Http2ErrorCode.ENHANCE_YOUR_CALM, kind + " queue exhausted");
+    }
+    for (ControlIntent intent : controls) {
+      if (intent.claim(kind)) return intent;
+    }
+    counter.decrementAndGet();
+    throw Http2Exception.of(Http2ErrorCode.ENHANCE_YOUR_CALM, "control-frame queue exhausted");
+  }
+
+  private void release(ControlIntent intent) {
+    counter(intent.kind).decrementAndGet();
+    intent.release();
+  }
+
+  private AtomicInteger counter(ControlKind kind) {
+    return kind == ControlKind.PING ? pingResponses : settingsAcks;
+  }
+
+  private static int limit(ControlKind kind) {
+    return kind == ControlKind.PING
+        ? Http2Limits.MAX_PING_QUEUE_DEPTH
+        : Http2Limits.MAX_SETTINGS_ACK_QUEUE_DEPTH;
+  }
+
+  enum ControlKind {
+    PING,
+    SETTINGS_OR_OTHER
+  }
+
+  static final class ControlIntent implements WriteIntent {
+    private final Http2ConnectionScratch owner;
+    private final byte[] buffer;
+    private final AtomicBoolean claimed = new AtomicBoolean();
+    private volatile WriteIntent next;
+    private ControlKind kind;
+    private int length;
+
+    private ControlIntent(Http2ConnectionScratch owner, int capacity) {
+      this.owner = owner;
+      this.buffer = new byte[capacity];
+    }
+
+    private boolean claim(ControlKind kind) {
+      if (!claimed.compareAndSet(false, true)) return false;
+      this.kind = kind;
+      this.length = 0;
+      this.next = null;
+      return true;
+    }
+
+    void copy(byte[] source) {
+      System.arraycopy(source, 0, buffer, 0, source.length);
+      length = source.length;
+    }
+
+    void frame(FrameType type, int flags, int streamId, byte[] payload, int off, int len) {
+      if (9 + len > buffer.length) {
+        throw new IllegalArgumentException("control frame exceeds scratch capacity");
+      }
+      buffer[0] = (byte) (len >>> 16);
+      buffer[1] = (byte) (len >>> 8);
+      buffer[2] = (byte) len;
+      buffer[3] = (byte) type.code();
+      buffer[4] = (byte) flags;
+      writeUInt31(buffer, 5, streamId);
+      System.arraycopy(payload, off, buffer, 9, len);
+      length = 9 + len;
+    }
+
+    void goAway(int lastStreamId, Http2ErrorCode error, String debug) {
+      int debugLength =
+          Math.min(
+              debug.length(),
+              Math.min(Http2Limits.MAX_GOAWAY_DEBUG_DATA_LENGTH, buffer.length - 17));
+      int payloadLength = 8 + debugLength;
+      buffer[0] = 0;
+      buffer[1] = 0;
+      buffer[2] = (byte) payloadLength;
+      buffer[3] = (byte) FrameType.GOAWAY.code();
+      buffer[4] = 0;
+      writeUInt31(buffer, 5, 0);
+      writeUInt31(buffer, 9, lastStreamId);
+      writeUInt32(buffer, 13, error.code());
+      for (int i = 0; i < debugLength; i++) buffer[17 + i] = (byte) debug.charAt(i);
+      length = 17 + debugLength;
+    }
+
+    void windowUpdate(int streamId, int increment) {
+      buffer[0] = 0;
+      buffer[1] = 0;
+      buffer[2] = 4;
+      buffer[3] = (byte) FrameType.WINDOW_UPDATE.code();
+      buffer[4] = 0;
+      writeUInt31(buffer, 5, streamId);
+      writeUInt31(buffer, 9, increment);
+      length = 13;
+    }
+
+    private static void writeUInt31(byte[] target, int off, int value) {
+      writeUInt32(target, off, value & 0x7FFF_FFFF);
+    }
+
+    private static void writeUInt32(byte[] target, int off, int value) {
+      target[off] = (byte) (value >>> 24);
+      target[off + 1] = (byte) (value >>> 16);
+      target[off + 2] = (byte) (value >>> 8);
+      target[off + 3] = (byte) value;
+    }
+
+    @Override
+    public byte[] buffer() {
+      return buffer;
+    }
+
+    @Override
+    public int offset() {
+      return 0;
+    }
+
+    @Override
+    public int length() {
+      return length;
+    }
+
+    @Override
+    public WriteIntent mpscNext() {
+      return next;
+    }
+
+    @Override
+    public void setMpscNext(WriteIntent next) {
+      this.next = next;
+    }
+
+    @Override
+    public void completed() {
+      owner.release(this);
+    }
+
+    private void release() {
+      next = null;
+      claimed.set(false);
+    }
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2ErrorCode.java

@@ -0,0 +1,94 @@
+package dev.relism.flash.http2;
+
+/**
+ * The 14 HTTP/2 error codes defined by RFC 9113 §7.
+ *
+ * <p>Each constant carries its 4-byte big-endian wire encoding, precomputed once at class
+ * load (RFC 9113 §6.4 {@code RST_STREAM} and §6.8 {@code GOAWAY} both carry the error code as
+ * a raw 32-bit field — there is no framing around it to build). Callers write
+ *
+ * <p>{@code Http2ErrorCode} is used to reject a peer <em>and</em> to interpret what a peer
+ * sends us: {@link #fromCode(int)} decodes a received 32-bit value. RFC 9113 does not reserve
+ * unknown codes for future use in a way that requires us to accept them silently as one of the
+ * 14 — an endpoint that receives an error code it does not recognise treats it as
+ * {@code INTERNAL_ERROR}-equivalent for logging purposes; {@link #fromCode(int)} returns
+ * {@code null} for that case and callers log the raw integer rather than guessing a mapping.
+ */
+public enum Http2ErrorCode {
+
+    /** Graceful shutdown or successful completion; not an error. RFC 9113 §7. */
+    NO_ERROR(0x00),
+    /** The peer violated the protocol in a way not covered by a more specific code. */
+    PROTOCOL_ERROR(0x01),
+    /** Unexpected internal condition on our side (e.g. an uncaught exception on the demux loop). */
+    INTERNAL_ERROR(0x02),
+    /** A flow-control window was violated: overflow past 2^31-1, or a peer exceeded its window. */
+    FLOW_CONTROL_ERROR(0x03),
+    /** The peer did not acknowledge our SETTINGS within {@code SETTINGS_ACK_TIMEOUT_MS}. */
+    SETTINGS_TIMEOUT(0x04),
+    /** A frame was received for a stream that is already closed. */
+    STREAM_CLOSED(0x05),
+    /** A frame's length did not match what its type requires (RFC 9113 §4.2, per-type rules). */
+    FRAME_SIZE_ERROR(0x06),
+    /** The stream was refused before any processing; safe for the client to retry elsewhere. */
+    REFUSED_STREAM(0x07),
+    /** Used by clients to cancel a stream; Flash never sends it, only receives it. */
+    CANCEL(0x08),
+    /** An HPACK decoding failure. Terminates the connection because the dynamic table state is lost. */
+    COMPRESSION_ERROR(0x09),
+    /** A CONNECT-tunnelled stream failed. */
+    CONNECT_ERROR(0x0a),
+    /** The peer is generating excessive load (rate-limit rejection: Rapid Reset, PING/SETTINGS floods). */
+    ENHANCE_YOUR_CALM(0x0b),
+    /** The negotiated TLS parameters fall below RFC 9113 §9.2's minimum security requirements. */
+    INADEQUATE_SECURITY(0x0c),
+    /** Defined by RFC 9113 for HTTP/1.1-only resources; Flash serves everything over h2, so unused. */
+    HTTP_1_1_REQUIRED(0x0d);
+
+    private static final Http2ErrorCode[] BY_CODE = new Http2ErrorCode[values().length];
+
+    static {
+        for (Http2ErrorCode c : values()) {
+            BY_CODE[c.code] = c;
+        }
+    }
+
+    private final int code;
+    private final byte[] bytes;
+
+    Http2ErrorCode(int code) {
+        this.code = code;
+        this.bytes = new byte[]{
+                (byte) (code >>> 24),
+                (byte) (code >>> 16),
+                (byte) (code >>> 8),
+                (byte) code
+        };
+    }
+
+    /** The numeric error code as it appears on the wire. */
+    public int code() {
+        return code;
+    }
+
+    /**
+     * The pre-encoded 4-byte big-endian wire form. Safe to write directly into a
+     * {@code RST_STREAM} or {@code GOAWAY} payload with a single {@code System.arraycopy} —
+     * never allocated or formatted per use.
+     */
+    public byte[] bytes() {
+        return bytes;
+    }
+
+    /**
+     * Decodes a 32-bit error code received from a peer. Returns {@code null} for a value
+     * outside the 14 defined codes; the caller should log the raw integer rather than assume
+     * a mapping, since RFC 9113 permits future extension codes we do not yet know about.
+     */
+    public static Http2ErrorCode fromCode(int code) {
+        if (code >= 0 && code < BY_CODE.length) {
+            return BY_CODE[code];
+        }
+        return null;
+    }
+}

flash/src/main/java/dev/relism/flash/http2/Http2Exception.java

@@ -0,0 +1,70 @@
+package dev.relism.flash.http2;
+
+/**
+ * A <b>connection-level</b> HTTP/2 error. Thrown anywhere a peer's frame, HPACK block, or
+ * SETTINGS value violates the protocol in a way that leaves the connection's state (the HPACK
+ * dynamic table, a flow-control window, the stream table) unrecoverable.
+ *
+ * single site: it sends {@code GOAWAY} with {@link #errorCode()} and closes the connection.
+ * Compare {@link Http2StreamException}, whose scope is one stream and which results in
+ * {@code RST_STREAM} while the connection survives.
+ *
+ * <p>Deliberately does <b>not</b> extend {@link java.io.IOException}: the connection loop
+ * distinguishes a protocol violation (a decision Flash made about the peer's bytes) from a
+ * socket failure (the peer went away) by catching these as unrelated types. Conflating them
+ * would make it possible to accidentally treat a hostile peer's malformed frame as a harmless
+ * disconnect, or vice versa.
+ *
+ * <h3>Why stack traces are disabled</h3>
+ * This exception is thrown on the connection's hot rejection path — a single malformed byte
+ * from a hostile or buggy peer can trigger it, and under a scripted attack that can happen many
+ * times per second across many connections. JVM stack trace capture ({@code fillInStackTrace})
+ * is by far the most expensive part of constructing a {@code Throwable}, and it buys nothing
+ * here: the call site is exactly where {@code errorCode()} says it is, and the debug message
+ * already names the specific violation. The 4-argument {@link RuntimeException} constructor
+ * disables both suppression and stack-trace writing.
+ *
+ * <h3>Preallocated singletons</h3>
+ * For the common, message-less rejections (frame validation failures, HPACK structural errors)
+ * this class exposes shared singleton instances. Reusing one instance across threads and across
+ * many throws is safe <em>only because</em> the instance carries no per-throw mutable state and
+ * writable-stack-trace is disabled — nothing about a throw mutates the exception object.
+ */
+public final class Http2Exception extends RuntimeException {
+
+    private final Http2ErrorCode errorCode;
+
+    private Http2Exception(Http2ErrorCode errorCode, String message) {
+        super(message, null, false, false);
+        this.errorCode = errorCode;
+    }
+
+    /** The RFC 9113 §7 error code to send in the {@code GOAWAY} frame. */
+    public Http2ErrorCode errorCode() {
+        return errorCode;
+    }
+
+    /**
+     * Builds a connection error carrying a caller-supplied debug message. Allocates a new
+     * message carries information specific to this occurrence (e.g. the offending stream id or
+     * a decoded value); use one of the preallocated singletons below when it does not.
+     */
+    public static Http2Exception of(Http2ErrorCode code, String message) {
+        return new Http2Exception(code, message);
+    }
+
+    // ── Preallocated, message-less singletons for the hot rejection paths ──────────────────
+
+    public static final Http2Exception PROTOCOL_ERROR =
+            new Http2Exception(Http2ErrorCode.PROTOCOL_ERROR, "protocol error");
+    public static final Http2Exception FRAME_SIZE_ERROR =
+            new Http2Exception(Http2ErrorCode.FRAME_SIZE_ERROR, "frame size error");
+    public static final Http2Exception FLOW_CONTROL_ERROR =
+            new Http2Exception(Http2ErrorCode.FLOW_CONTROL_ERROR, "flow control error");
+    public static final Http2Exception COMPRESSION_ERROR =
+            new Http2Exception(Http2ErrorCode.COMPRESSION_ERROR, "compression error");
+    public static final Http2Exception INTERNAL_ERROR =
+            new Http2Exception(Http2ErrorCode.INTERNAL_ERROR, "internal error");
+    public static final Http2Exception SETTINGS_TIMEOUT =
+            new Http2Exception(Http2ErrorCode.SETTINGS_TIMEOUT, "settings ack timeout");
+}

flash/src/main/java/dev/relism/flash/http2/Http2HeaderBlockDecoder.java

@@ -0,0 +1,115 @@
+package dev.relism.flash.http2;
+
+import dev.relism.flash.bytes.Pairs;
+import dev.relism.flash.http2.frame.FrameFlags;
+import dev.relism.flash.http2.frame.FrameHeader;
+import dev.relism.flash.http2.frame.FrameType;
+import dev.relism.flash.http2.frame.Padding;
+import dev.relism.flash.http2.hpack.ContinuationAssembler;
+import dev.relism.flash.http2.hpack.HeaderListSizeException;
+import dev.relism.flash.http2.hpack.HeaderSink;
+import dev.relism.flash.http2.hpack.HpackDecoder;
+
+/** Composes frame fragment extraction, CONTINUATION assembly and HPACK decoding. */
+final class Http2HeaderBlockDecoder {
+  private static final int PRIORITY_FIELDS_LENGTH = 5;
+
+  private final ContinuationAssembler assembler = new ContinuationAssembler();
+  private final HpackDecoder decoder = new HpackDecoder();
+  private final long assemblyTimeoutNanos;
+  private boolean endStream;
+  private long assemblyStartedNanos;
+
+  Http2HeaderBlockDecoder() {
+    this(Http2Limits.HEADER_BLOCK_ASSEMBLY_TIMEOUT_MS);
+  }
+
+  Http2HeaderBlockDecoder(long assemblyTimeoutMillis) {
+    if (assemblyTimeoutMillis <= 0) throw new IllegalArgumentException("non-positive timeout");
+    assemblyTimeoutNanos = assemblyTimeoutMillis * 1_000_000L;
+  }
+
+  boolean insideHeaderBlock() {
+    return assembler.isActive();
+  }
+
+  /** Accepts HEADERS or CONTINUATION and returns true when a complete block was decoded. */
+  boolean accept(FrameHeader frame, HeaderSink sink) {
+    checkTimeout();
+    if (assembler.isActive() && frame.type() != FrameType.CONTINUATION) {
+      throw Http2Exception.PROTOCOL_ERROR;
+    }
+    if (frame.type() == FrameType.HEADERS) {
+      begin(frame);
+    } else if (frame.type() == FrameType.CONTINUATION) {
+      assembler.continuation(
+          frame.streamId(),
+          frame.buffer(),
+          frame.payloadOffset(),
+          frame.length(),
+          FrameFlags.isEndHeaders(frame.flags()));
+    } else {
+      return false;
+    }
+    if (!assembler.isComplete()) return false;
+
+    try {
+      decoder.decode(assembler.buffer(), 0, assembler.length(), sink);
+    } catch (HeaderListSizeException tooLarge) {
+      int streamId = assembler.streamId();
+      assembler.reset();
+      throw new Http2StreamException(
+          streamId, Http2ErrorCode.ENHANCE_YOUR_CALM, tooLarge.getMessage());
+    }
+    assembler.reset();
+    assemblyStartedNanos = 0;
+    return true;
+  }
+
+  void checkTimeout() {
+    if (assembler.isActive()
+        && System.nanoTime() - assemblyStartedNanos >= assemblyTimeoutNanos) {
+      assembler.reset();
+      assemblyStartedNanos = 0;
+      throw Http2Exception.of(Http2ErrorCode.ENHANCE_YOUR_CALM,
+          "header block assembly timeout");
+    }
+  }
+
+  boolean endStream() {
+    return endStream;
+  }
+
+  private void begin(FrameHeader frame) {
+    assemblyStartedNanos = System.nanoTime();
+    endStream = FrameFlags.isEndStream(frame.flags());
+    long unpadded =
+        Padding.unpad(
+            frame.buffer(),
+            frame.payloadOffset(),
+            frame.length(),
+            FrameFlags.isPadded(frame.flags()));
+    int fragmentOffset = Pairs.hi(unpadded);
+    int fragmentLength = Pairs.lo(unpadded);
+    if (FrameFlags.hasPriority(frame.flags())) {
+      if (fragmentLength < PRIORITY_FIELDS_LENGTH) throw Http2Exception.FRAME_SIZE_ERROR;
+      int dependency =
+          ((frame.buffer()[fragmentOffset] & 0x7f) << 24)
+              | ((frame.buffer()[fragmentOffset + 1] & 0xff) << 16)
+              | ((frame.buffer()[fragmentOffset + 2] & 0xff) << 8)
+              | (frame.buffer()[fragmentOffset + 3] & 0xff);
+      if (dependency == frame.streamId()) {
+        throw new Http2StreamException(
+            frame.streamId(), Http2ErrorCode.PROTOCOL_ERROR, "stream depends on itself");
+      }
+      fragmentOffset += PRIORITY_FIELDS_LENGTH;
+      fragmentLength -= PRIORITY_FIELDS_LENGTH;
+    }
+    assembler.begin(
+        frame.streamId(),
+        frame.buffer(),
+        fragmentOffset,
+        fragmentLength,
+        FrameFlags.isEndHeaders(frame.flags()));
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2Limits.java

@@ -0,0 +1,219 @@
+package dev.relism.flash.http2;
+
+/**
+ * Every bound the HTTP/2 implementation enforces against a peer's input, in one place.
+ *
+ * <p>Every wire-derived length, index, count, or size is checked against a named constant here —
+ * never against an ad-hoc literal, and never by letting the underlying array or buffer throw on
+ * overrun. Each field's Javadoc names the specific attack or resource it bounds and, where one
+ * exists, the CVE.
+ *
+ * <p>These are compile-time defaults, not runtime configuration. A limit becomes configurable only
+ * when the operational need and its safe range are established.
+ *
+ * <p>Each field is introduced with the feature that enforces it; this class contains no unused
+ * placeholders.
+ */
+public final class Http2Limits {
+
+  private Http2Limits() {}
+
+  /**
+   * Maximum number of streams a single connection may have open concurrently. Advertised to the
+   * peer as {@code SETTINGS_MAX_CONCURRENT_STREAMS}. Bounds per-connection memory (each open stream
+   * owns a per-stream HPACK arena and request/response state) against a peer that simply opens
+   * streams and never closes them.
+   */
+  public static final int MAX_CONCURRENT_STREAMS = 64;
+
+  /**
+   * The largest frame payload we accept without the peer first raising it via our own {@code
+   * SETTINGS_MAX_FRAME_SIZE}. RFC 9113 §4.2 fixes the protocol default at 16384 and requires any
+   * advertised value to stay within {@code 16384..16777215}. Bounds the memory a single frame read
+   * can force us to hold.
+   */
+  public static final int MAX_FRAME_SIZE_LOCAL = 16_384;
+
+  /**
+   * Maximum total size (name + value + 32 per RFC 7541 §4.1's accounting, summed over every header)
+   * of a decoded header list. Advertised as {@code SETTINGS_MAX_HEADER_LIST_SIZE} (RFC 9113
+   * §6.5.2). This is the primary defence against an HPACK bomb: a small compressed block that
+   * references dynamic-table entries to expand into an enormous header list.
+   */
+  public static final int MAX_HEADER_LIST_SIZE = 32_768;
+
+  /**
+   * Maximum number of CONTINUATION frames accepted for a single header block before the connection
+   * is torn down. Defence against CVE-2024-27316 (the "HTTP/2 CONTINUATION Flood"): a peer that
+   * never sets {@code END_HEADERS} can otherwise force unbounded decode/reassembly work per header
+   * block.
+   */
+  public static final int MAX_CONTINUATION_FRAMES_PER_BLOCK = 8;
+
+  /**
+   * Maximum number of {@code RST_STREAM} frames accepted from the peer within {@link
+   * #RESET_RATE_INTERVAL_MS}. Defence against CVE-2023-44487 ("HTTP/2 Rapid Reset"): opening a
+   * stream and immediately resetting it does not count against {@link #MAX_CONCURRENT_STREAMS}, so
+   * without a rate bound a peer can force unbounded per-stream setup/teardown work at effectively
+   * unlimited concurrency.
+   */
+  public static final int MAX_RESET_STREAMS_PER_INTERVAL = 200;
+
+  /**
+   * The rolling window (milliseconds) over which {@link #MAX_RESET_STREAMS_PER_INTERVAL} is
+   * measured.
+   */
+  public static final long RESET_RATE_INTERVAL_MS = 10_000;
+
+  /**
+   * Maximum number of new streams accepted from the peer within {@link #RESET_RATE_INTERVAL_MS}. A
+   * companion bound to {@link #MAX_RESET_STREAMS_PER_INTERVAL}: Rapid Reset defences that only
+   * count resets can still be bypassed by a peer that creates streams fast enough that the reset
+   * counter never saturates within any single window boundary.
+   *
+   * <p>Matches {@link #MAX_STREAMS_PER_CONNECTION}'s lifetime budget by design: a connection may
+   * not create more streams in one rolling burst window than it is ever allowed to create in its
+   * whole lifetime. An earlier value of 400 (40/s) measured the RST_STREAM flood attack this bound
+   * exists for, but also rejected ordinary high-concurrency multiplexed clients well below the
+   * throughput a hardened server is expected to sustain — h2load's default light-load pattern (10
+   * connections, 10 concurrent streams each) alone drives multiple thousands of legitimate stream
+   * creations per connection per second on a fast peer, which 400/10s cannot distinguish from
+   * abuse. The RST_STREAM-rate counter above measures the actual CVE-2023-44487 signature (resets,
+   * not creates); this bound only needs to catch a peer creating streams fast enough to dodge that
+   * counter, which a much higher ceiling still does.
+   */
+  public static final int MAX_STREAMS_CREATED_PER_INTERVAL = 100_000;
+
+  /** Maximum SETTINGS frames accepted within one abuse-rate interval. */
+  public static final int MAX_SETTINGS_PER_INTERVAL = 100;
+
+  /** Maximum non-acknowledgement PING frames accepted within one abuse-rate interval. */
+  public static final int MAX_PINGS_PER_INTERVAL = 200;
+
+  /**
+   * Aggregate bound for frames that consume parsing work without carrying application data:
+   * PRIORITY, WINDOW_UPDATE, empty DATA and unknown extension frames.
+   */
+  public static final int MAX_USELESS_FRAMES_PER_INTERVAL = 10_000;
+
+  /** Default total-stream budget for one connection; zero disables the budget. */
+  public static final long MAX_STREAMS_PER_CONNECTION = 100_000;
+
+  /** Default wire-byte budget for one connection; zero disables the budget. */
+  public static final long MAX_BYTES_PER_CONNECTION = 0;
+
+  /** Default connection lifetime budget in milliseconds; zero disables the budget. */
+  public static final long MAX_CONNECTION_LIFETIME_MS = 0;
+
+  /**
+   * Maximum number of SETTINGS parameter entries accepted in a single SETTINGS frame. A SETTINGS
+   * frame is already bounded in byte length by {@link #MAX_FRAME_SIZE_LOCAL} (each entry is 6
+   * bytes), but an explicit entry-count bound keeps the per-entry validation loop itself cheap to
+   * reason about and gives a distinct, loud rejection reason.
+   */
+  public static final int MAX_SETTINGS_ENTRIES_PER_FRAME = 64;
+
+  /** Maximum number of locally-sent SETTINGS frames awaiting acknowledgement. */
+  public static final int MAX_OUTSTANDING_LOCAL_SETTINGS = 8;
+
+  /** Maximum time allowed for the peer to acknowledge a locally-sent SETTINGS frame. */
+  public static final long SETTINGS_ACK_TIMEOUT_MS = 10_000;
+
+  /**
+   * Maximum number of SETTINGS acknowledgements waiting behind a blocked socket writer. This
+   * prevents a peer from turning a stream of empty SETTINGS frames into an unbounded queue of
+   * mandatory responses.
+   */
+  public static final int MAX_SETTINGS_ACK_QUEUE_DEPTH = 64;
+
+  /** Maximum diagnostic bytes included in an outbound GOAWAY frame. */
+  public static final int MAX_GOAWAY_DEBUG_DATA_LENGTH = 128;
+
+  /**
+   * Maximum number of outstanding (unanswered) PING responses queued for the writer. A PING flood
+   * forces a PONG per PING; without a bound, a peer that reads its own responses slowly can make us
+   * buffer unbounded PONG frames.
+   */
+  public static final int MAX_PING_QUEUE_DEPTH = 64;
+
+  /**
+   * Maximum number of zero-length DATA frames accepted per stream. Zero-length DATA consumes no
+   * flow-control window, so window accounting does not bound it — without this limit a peer can
+   * force unbounded per-frame dispatch/validation CPU work at zero cost to itself.
+   */
+  public static final int MAX_EMPTY_DATA_FRAMES_PER_STREAM = 1_000;
+
+  /** Largest request body retained contiguously before dispatching its handler. */
+  public static final int INLINE_BODY_THRESHOLD = 64 * 1024;
+
+  /** Hard limit for request body bytes accepted on one stream. */
+  public static final int MAX_REQUEST_BODY_SIZE = 100 * 1024 * 1024;
+
+  /** Number of frame-sized buffers available to streaming request bodies on one connection. */
+  public static final int DATA_BUFFER_POOL_SIZE = 64;
+
+  /**
+   * The value of {@code SETTINGS_INITIAL_WINDOW_SIZE} Flash advertises for every new stream:
+   * deliberately large (1 MiB, versus the RFC default of 65535) so that a normal-sized
+   */
+  public static final int INITIAL_WINDOW_SIZE_LOCAL = 1_048_576;
+
+  /**
+   * The connection-level flow-control window Flash advertises. Sized above {@link
+   * #INITIAL_WINDOW_SIZE_LOCAL} so a single active stream is never bottlenecked by the connection
+   * window before its own stream window, but well below {@code MAX_CONCURRENT_STREAMS *
+   * INITIAL_WINDOW_SIZE_LOCAL} — real traffic is never all streams simultaneously saturating their
+   * windows, and sizing for that worst case would commit 100 MiB of receive window to every
+   * connection regardless of load.
+   */
+  public static final int CONNECTION_WINDOW_SIZE_LOCAL = 1_048_576;
+
+  /**
+   * The HPACK dynamic table size Flash's decoder honours, in bytes of RFC 7541 §4.1 accounting. RFC
+   * 7541's protocol default. The encoder never uses a dynamic table at all
+   */
+  public static final int HPACK_DYNAMIC_TABLE_SIZE_LOCAL = 4_096;
+
+  /**
+   * Maximum length, in decoded bytes, of a single HPACK string literal. Applied during Huffman
+   * decode as bytes are produced, not to the encoded length — a Huffman string can expand by
+   * roughly 8/5, so bounding only the encoded length would let a compact input still decode past
+   * this limit.
+   */
+  public static final int MAX_HPACK_STRING_LENGTH = 8_192;
+
+  /**
+   * Maximum time, in milliseconds, allowed between a HEADERS frame's arrival and the header block's
+   * completion (its {@code END_HEADERS} flag, possibly after CONTINUATION frames). A peer that
+   * starts a header block and then stalls indefinitely would otherwise hold the per-stream arena
+   * and the connection's HPACK assembly buffer forever.
+   */
+  public static final long HEADER_BLOCK_ASSEMBLY_TIMEOUT_MS = 10_000;
+
+  /**
+   * Maximum time, in milliseconds, a stream may remain open with no frame activity in either
+   * direction. Bounds resource pinning by a peer that opens a stream and then goes silent without
+   * closing it — the h2 equivalent of the h1 slowloris defence in {@code
+   * FlashConfiguration.idleKeepAliveTimeoutMs}.
+   */
+  public static final long STREAM_IDLE_TIMEOUT_MS = 60_000;
+
+  /**
+   * Maximum time, in milliseconds, {@code Http2FrameWriter} may spend blocked inside a single
+   * socket write. A blocking write is unavoidable when the kernel send buffer is full and the peer
+   * is not reading (that peer holds the connection's single writer lock for the duration — see
+   * {@code WRITER.md}), but it must not be unbounded: a peer that simply stops reading would
+   * otherwise let a single stalled connection wedge the writer forever. Enforced via a background
+   * reaper interrupting the blocked thread past the deadline, not {@code Socket#setSoTimeout} —
+   * that option bounds reads, not writes.
+   */
+  public static final long WRITE_TIMEOUT_MS = 30_000;
+
+  /**
+   * Maximum time, in milliseconds, {@code Http2FrameReader} may wait for a single frame's
+   * header and payload to fully arrive. Bounds the same slowloris-shaped hazard:
+   * without it, a peer that sends 9 header bytes and then never sends the declared payload
+   * would hold this connection's frame reader waiting forever.
+   */
+  public static final long FRAME_READ_TIMEOUT_MS = 20_000;
+}

flash/src/main/java/dev/relism/flash/http2/Http2Preface.java

@@ -0,0 +1,87 @@
+package dev.relism.flash.http2;
+
+import dev.relism.flash.bytes.ByteWriter;
+import dev.relism.flash.http2.frame.FrameFlags;
+import dev.relism.flash.http2.frame.FrameType;
+import dev.relism.flash.http2.frame.FrameWriteBuffer;
+import java.nio.charset.StandardCharsets;
+
+/** Byte-exact client preface and immutable server startup frames, compiled once at class load. */
+public final class Http2Preface {
+  private static final byte[] CLIENT_PREFACE =
+      "PRI * HTTP/2.0\r\n\r\nSM\r\n\r\n".getBytes(StandardCharsets.US_ASCII);
+  private static final byte[] SERVER_SETTINGS = buildServerSettings();
+  private static final byte[] SETTINGS_ACK = frame(FrameType.SETTINGS, FrameFlags.ACK, 0, 0);
+  private static final byte[] INITIAL_CONNECTION_WINDOW = buildInitialConnectionWindow();
+
+  private Http2Preface() {}
+
+  /** Immutable client connection preface bytes. Callers must not modify the returned array. */
+  public static byte[] clientPreface() {
+    return CLIENT_PREFACE;
+  }
+
+  static int clientPrefaceLength() {
+    return CLIENT_PREFACE.length;
+  }
+
+  static boolean matchesClientPreface(byte[] candidate) {
+    if (candidate.length != CLIENT_PREFACE.length) return false;
+    int different = 0;
+    for (int i = 0; i < CLIENT_PREFACE.length; i++) {
+      different |= candidate[i] ^ CLIENT_PREFACE[i];
+    }
+    return different == 0;
+  }
+
+  static byte[] serverSettings() {
+    return SERVER_SETTINGS;
+  }
+
+  static byte[] settingsAck() {
+    return SETTINGS_ACK;
+  }
+
+  static byte[] initialConnectionWindow() {
+    return INITIAL_CONNECTION_WINDOW;
+  }
+
+  private static byte[] buildServerSettings() {
+    ByteWriter bytes = new ByteWriter(64);
+    FrameWriteBuffer frame = new FrameWriteBuffer(bytes);
+    frame.beginFrame(FrameType.SETTINGS, 0, 0);
+    setting(bytes, Http2Settings.HEADER_TABLE_SIZE, Http2Limits.HPACK_DYNAMIC_TABLE_SIZE_LOCAL);
+    setting(bytes, Http2Settings.ENABLE_PUSH, 0);
+    setting(bytes, Http2Settings.MAX_CONCURRENT_STREAMS, Http2Limits.MAX_CONCURRENT_STREAMS);
+    setting(bytes, Http2Settings.INITIAL_WINDOW_SIZE, Http2Limits.INITIAL_WINDOW_SIZE_LOCAL);
+    setting(bytes, Http2Settings.MAX_HEADER_LIST_SIZE, Http2Limits.MAX_HEADER_LIST_SIZE);
+    setting(bytes, Http2Settings.ENABLE_CONNECT_PROTOCOL, 1);
+    frame.endFrame();
+    return copy(bytes);
+  }
+
+  private static byte[] buildInitialConnectionWindow() {
+    int increment = Http2Limits.CONNECTION_WINDOW_SIZE_LOCAL - 65_535;
+    return frame(FrameType.WINDOW_UPDATE, 0, 0, increment);
+  }
+
+  private static byte[] frame(FrameType type, int flags, int streamId, int payload) {
+    ByteWriter bytes = new ByteWriter(16);
+    FrameWriteBuffer frame = new FrameWriteBuffer(bytes);
+    frame.beginFrame(type, flags, streamId);
+    if (type == FrameType.WINDOW_UPDATE) bytes.writeUInt31(payload);
+    frame.endFrame();
+    return copy(bytes);
+  }
+
+  private static void setting(ByteWriter bytes, int id, int value) {
+    bytes.writeUInt16(id);
+    bytes.writeUInt32(value);
+  }
+
+  private static byte[] copy(ByteWriter bytes) {
+    byte[] result = new byte[bytes.length()];
+    System.arraycopy(bytes.array(), 0, result, 0, result.length);
+    return result;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2Settings.java

@@ -0,0 +1,149 @@
+package dev.relism.flash.http2;
+
+/**
+ * The peer's HTTP/2 SETTINGS state. A received payload is validated completely before any value is
+ * applied, so a malformed parameter cannot leave a partially-updated connection.
+ */
+public final class Http2Settings {
+  public static final int HEADER_TABLE_SIZE = 0x1;
+  public static final int ENABLE_PUSH = 0x2;
+  public static final int MAX_CONCURRENT_STREAMS = 0x3;
+  public static final int INITIAL_WINDOW_SIZE = 0x4;
+  public static final int MAX_FRAME_SIZE = 0x5;
+  public static final int MAX_HEADER_LIST_SIZE = 0x6;
+  public static final int ENABLE_CONNECT_PROTOCOL = 0x8;
+
+  public static final int DEFAULT_HEADER_TABLE_SIZE = 4_096;
+  public static final int DEFAULT_INITIAL_WINDOW_SIZE = 65_535;
+  public static final int DEFAULT_MAX_FRAME_SIZE = 16_384;
+
+  /**
+   * Applies an INITIAL_WINDOW_SIZE delta to every open stream. Implementations must validate all
+   * resulting windows before changing any of them; negative results are valid, while a result above
+   * {@link Integer#MAX_VALUE} is a connection FLOW_CONTROL_ERROR.
+   */
+  @FunctionalInterface
+  public interface StreamWindowUpdater {
+    void applyInitialWindowDelta(int delta);
+  }
+
+  private int headerTableSize = DEFAULT_HEADER_TABLE_SIZE;
+  private boolean pushEnabled = true;
+  private long maxConcurrentStreams = 0xFFFF_FFFFL;
+  private int initialWindowSize = DEFAULT_INITIAL_WINDOW_SIZE;
+  private int maxFrameSize = DEFAULT_MAX_FRAME_SIZE;
+  private long maxHeaderListSize = 0xFFFF_FFFFL;
+
+  /** Validates and applies one SETTINGS payload. Unknown identifiers are ignored. */
+  public void apply(byte[] payload, int off, int len, StreamWindowUpdater streamWindows) {
+    if (len % 6 != 0) throw Http2Exception.FRAME_SIZE_ERROR;
+    int entries = len / 6;
+    if (entries > Http2Limits.MAX_SETTINGS_ENTRIES_PER_FRAME) {
+      throw Http2Exception.of(
+          Http2ErrorCode.ENHANCE_YOUR_CALM, "too many SETTINGS entries: " + entries);
+    }
+    checkRange(payload, off, len);
+
+    int nextHeaderTableSize = headerTableSize;
+    boolean nextPushEnabled = pushEnabled;
+    long nextMaxConcurrentStreams = maxConcurrentStreams;
+    int nextInitialWindowSize = initialWindowSize;
+    int nextMaxFrameSize = maxFrameSize;
+    long nextMaxHeaderListSize = maxHeaderListSize;
+    for (int pos = off; pos < off + len; pos += 6) {
+      int id = readUInt16(payload, pos);
+      long value = readUInt32(payload, pos + 2);
+      validate(id, value);
+      switch (id) {
+        case HEADER_TABLE_SIZE ->
+            nextHeaderTableSize = (int) Math.min(value, Http2Limits.HPACK_DYNAMIC_TABLE_SIZE_LOCAL);
+        case ENABLE_PUSH -> nextPushEnabled = value == 1;
+        case MAX_CONCURRENT_STREAMS -> nextMaxConcurrentStreams = value;
+        case INITIAL_WINDOW_SIZE -> nextInitialWindowSize = (int) value;
+        case MAX_FRAME_SIZE -> nextMaxFrameSize = (int) value;
+        case MAX_HEADER_LIST_SIZE -> nextMaxHeaderListSize = value;
+        default -> {
+          // RFC 9113 §6.5.2: ignore unknown settings.
+        }
+      }
+    }
+
+    streamWindows.applyInitialWindowDelta(nextInitialWindowSize - initialWindowSize);
+    headerTableSize = nextHeaderTableSize;
+    pushEnabled = nextPushEnabled;
+    maxConcurrentStreams = nextMaxConcurrentStreams;
+    initialWindowSize = nextInitialWindowSize;
+    maxFrameSize = nextMaxFrameSize;
+    maxHeaderListSize = nextMaxHeaderListSize;
+  }
+
+  private static void validate(int id, long value) {
+    switch (id) {
+      case ENABLE_PUSH, ENABLE_CONNECT_PROTOCOL -> {
+        if (value > 1) throw Http2Exception.PROTOCOL_ERROR;
+      }
+      case INITIAL_WINDOW_SIZE -> {
+        if (value > Integer.MAX_VALUE) throw Http2Exception.FLOW_CONTROL_ERROR;
+      }
+      case MAX_FRAME_SIZE -> {
+        if (value < 16_384 || value > 16_777_215) {
+          throw Http2Exception.PROTOCOL_ERROR;
+        }
+      }
+      default -> {
+        // HEADER_TABLE_SIZE, MAX_CONCURRENT_STREAMS and MAX_HEADER_LIST_SIZE accept
+        // every unsigned 32-bit value. Unknown identifiers are ignored by the RFC.
+      }
+    }
+  }
+
+  private static void checkRange(byte[] payload, int off, int len) {
+    if (off < 0 || len < 0 || off > payload.length - len) {
+      throw new IndexOutOfBoundsException("invalid SETTINGS payload range");
+    }
+  }
+
+  private static int readUInt16(byte[] buf, int off) {
+    return ((buf[off] & 0xFF) << 8) | (buf[off + 1] & 0xFF);
+  }
+
+  private static long readUInt32(byte[] buf, int off) {
+    return ((long) (buf[off] & 0xFF) << 24)
+        | ((long) (buf[off + 1] & 0xFF) << 16)
+        | ((long) (buf[off + 2] & 0xFF) << 8)
+        | (buf[off + 3] & 0xFFL);
+  }
+
+  public int headerTableSize() {
+    return headerTableSize;
+  }
+
+  public boolean pushEnabled() {
+    return pushEnabled;
+  }
+
+  public long maxConcurrentStreams() {
+    return maxConcurrentStreams;
+  }
+
+  public int initialWindowSize() {
+    return initialWindowSize;
+  }
+
+  public int maxFrameSize() {
+    return maxFrameSize;
+  }
+
+  public long maxHeaderListSize() {
+    return maxHeaderListSize;
+  }
+
+  void reset() {
+    headerTableSize = DEFAULT_HEADER_TABLE_SIZE;
+    pushEnabled = true;
+    maxConcurrentStreams = 0xFFFF_FFFFL;
+    initialWindowSize = DEFAULT_INITIAL_WINDOW_SIZE;
+    maxFrameSize = DEFAULT_MAX_FRAME_SIZE;
+    maxHeaderListSize = 0xFFFF_FFFFL;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2StreamDispatcher.java

@@ -0,0 +1,305 @@
+package dev.relism.flash.http2;
+
+import dev.relism.flash.http.ContentType;
+import dev.relism.flash.http.HttpMethod;
+import dev.relism.flash.http.HttpStatus;
+import dev.relism.flash.http2.frame.Http2FrameWriter;
+import dev.relism.flash.http2.message.Http2ResponseWriter;
+import dev.relism.flash.http2.stream.Http2FlowController;
+import dev.relism.flash.http2.stream.Http2Stream;
+import dev.relism.flash.http2.stream.Http2StreamState;
+import dev.relism.flash.http2.stream.Http2StreamTable;
+import dev.relism.flash.models.Request;
+import dev.relism.flash.models.RequestHandler;
+import dev.relism.flash.models.Response;
+import dev.relism.flash.models.ResponseStreamOutputStream;
+import dev.relism.flash.transport.ConnectionContext;
+import dev.relism.flash.websocket.WebSocketHandler;
+import dev.relism.flash.websocket.WebSocketLoop;
+import dev.relism.flash.websocket.WebSocketSession;
+import java.io.IOException;
+import java.util.concurrent.RejectedExecutionException;
+import lombok.extern.slf4j.Slf4j;
+
+/** Dispatches completed request streams without blocking the connection demultiplexer. */
+@Slf4j
+final class Http2StreamDispatcher implements Http2Stream.ResponseSink {
+  @FunctionalInterface
+  interface FailureSink {
+    void fail(int streamId, Http2ErrorCode errorCode) throws IOException;
+  }
+
+  private final ConnectionContext context;
+  private final Http2FrameWriter frameWriter;
+  private final Http2Settings peerSettings;
+  private final Http2StreamTable streams;
+  private final Http2FlowController flowController;
+  private final FailureSink failures;
+  private final Http2Stream[] resumeScratch =
+      new Http2Stream[Http2Limits.MAX_CONCURRENT_STREAMS];
+  private volatile boolean firstResponse = true;
+
+  Http2StreamDispatcher(
+      ConnectionContext context,
+      Http2FrameWriter frameWriter,
+      Http2Settings peerSettings,
+      Http2StreamTable streams,
+      Http2FlowController flowController,
+      FailureSink failures) {
+    this.context = context;
+    this.frameWriter = frameWriter;
+    this.peerSettings = peerSettings;
+    this.streams = streams;
+    this.flowController = flowController;
+    this.failures = failures;
+  }
+
+  void streamWindowUpdated(Http2Stream stream) {
+    scheduleResume(stream);
+  }
+
+  void connectionWindowUpdated() {
+    int count = streams.copyValues(resumeScratch);
+    for (int i = 0; i < count; i++) {
+      Http2Stream stream = resumeScratch[i];
+      resumeScratch[i] = null;
+      scheduleResume(stream);
+    }
+  }
+
+  private void scheduleResume(Http2Stream stream) {
+    if (!stream.responseStarted() || stream.cancelled()) return;
+    if (!stream.beginResponseBatch()) return;
+    stream.markResumeTask();
+    try {
+      context.executor().execute(stream);
+    } catch (RejectedExecutionException rejected) {
+      stream.endResponseBatch();
+      failAndRelease(stream, Http2ErrorCode.REFUSED_STREAM, rejected);
+    }
+  }
+
+  void dispatch(Http2Stream stream) {
+    if (stream.cancelled()) {
+      streams.release(stream);
+      return;
+    }
+    stream.markDispatched();
+    stream.responseSink(this);
+    try {
+      context.executor().execute(stream);
+    } catch (RejectedExecutionException rejected) {
+      failAndRelease(stream, Http2ErrorCode.REFUSED_STREAM, rejected);
+    }
+  }
+
+  @Override
+  public void handleRequest(Http2Stream stream) {
+    handle(stream);
+  }
+
+  private void handle(Http2Stream stream) {
+    stream.touch();
+    if (stream.cancelled()) {
+      streams.release(stream);
+      return;
+    }
+    try {
+      Request request = stream.assembleRequest(context.remoteAddress(), context.sslSocket());
+      Response pooled = stream.resetResponse();
+      Response response = pooled;
+      if (!Http2Authority.isServed(request.header("host"), request.sslSession())) {
+        response.status(HttpStatus.MISDIRECTED_REQUEST);
+        if (stream.websocketConnect()) response.type(ContentType.NONE).streaming(output -> {});
+      } else if (stream.websocketConnect()) {
+        WebSocketHandler handler =
+            context.wsRouter().route(request, stream.wsRouteScratch(context.wsRouter()));
+        response.type(ContentType.NONE);
+        if (handler == null) {
+          response.status(HttpStatus.NOT_FOUND).streaming(output -> {});
+        } else {
+          response.streaming(
+              output ->
+                  WebSocketLoop.run(
+                      new WebSocketSession(
+                          request.body().stream(),
+                          new ResponseStreamOutputStream(output),
+                          context.configuration().getWsFrameBufferSize(),
+                          request,
+                          false),
+                      handler));
+        }
+      } else {
+        Object routeScratch = stream.routeScratch(context.router());
+        RequestHandler handler = context.router().route(request, routeScratch);
+        if (handler == null) handler = context.router().getNotFoundHandler();
+        try {
+          Object result = handler.handle(request, response);
+          if (result instanceof Response returned) response = returned;
+          else if (result != null) response.setBody(result);
+        } catch (Exception handlerFailure) {
+          Object result =
+              context.router().getExceptionHandler().handle(handlerFailure, request, response);
+          if (result instanceof Response returned) response = returned;
+          else if (result != null) response.setBody(result);
+        }
+      }
+
+      boolean pushStreaming = response.isPushStreaming();
+      if (!pushStreaming) request.drain();
+      Http2ResponseWriter responseWriter = stream.responseWriter();
+      if (stream.cancelled()) {
+        request.recycle();
+        if (response == pooled) pooled.recycle();
+        streams.release(stream);
+        return;
+      }
+      boolean headRequest = request.method() == HttpMethod.HEAD;
+      int reserved;
+      int used;
+      synchronized (this) {
+        boolean tableUpdate = firstResponse;
+        reserved = flowController.reserveSend(stream, peerSettings.maxFrameSize());
+        used = 0;
+        try {
+          used =
+              responseWriter.startFlowControlled(
+                  response,
+                  stream.id(),
+                  headRequest,
+                  context.configuration().isSendDate(),
+                  true,
+                  context.configuration().isH2HuffmanDynamicValues(),
+                  tableUpdate,
+                  peerSettings.maxFrameSize(),
+                  peerSettings.maxHeaderListSize(),
+                  reserved);
+        } finally {
+          flowController.refundSend(stream, reserved - used);
+        }
+        firstResponse = false;
+      }
+      if (!pushStreaming) request.recycle();
+      stream.markResponseStarted();
+      applyBatchTransition(stream, responseWriter);
+      if (!stream.beginResponseBatch()) {
+        throw new IllegalStateException("response batch already in flight");
+      }
+      detachFinalBatch(stream, responseWriter);
+      frameWriter.write(responseWriter);
+    } catch (Exception failure) {
+      failAndRelease(stream, Http2ErrorCode.INTERNAL_ERROR, failure);
+    }
+  }
+
+  private void tryResumeResponse(Http2Stream stream) {
+    stream.touch();
+    if (stream.cancelled()) {
+      stream.endResponseBatch();
+      streams.release(stream);
+      return;
+    }
+    Http2ResponseWriter responseWriter = stream.responseWriter();
+    int streamId = stream.id();
+    if (responseWriter.finished()) {
+      stream.endResponseBatch();
+      if (stream.state() == Http2StreamState.CLOSED) {
+        streams.retire(stream, streamId);
+      }
+      return;
+    }
+    int reserved = flowController.reserveSend(stream, peerSettings.maxFrameSize());
+    if (reserved == 0) {
+      stream.endResponseBatch();
+      return;
+    }
+    try {
+      int used = 0;
+      try {
+        used = responseWriter.resume(peerSettings.maxFrameSize(), reserved);
+      } finally {
+        flowController.refundSend(stream, reserved - used);
+      }
+      applyBatchTransition(stream, responseWriter);
+      detachFinalBatch(stream, responseWriter);
+      frameWriter.write(responseWriter);
+    } catch (Exception failure) {
+      stream.endResponseBatch();
+      failAndRelease(stream, Http2ErrorCode.INTERNAL_ERROR, failure);
+    }
+  }
+
+  @Override
+  public void resumeResponse(Http2Stream stream) {
+    tryResumeResponse(stream);
+  }
+
+  private static void applyBatchTransition(
+      Http2Stream stream, Http2ResponseWriter responseWriter) {
+    if (responseWriter.headersInBatch()) {
+      if (responseWriter.finished() && responseWriter.dataBytesInBatch() == 0
+          && !responseWriter.trailerHeadersInBatch()) {
+        stream.transition(Http2StreamState.Event.SEND_HEADERS_ES);
+        return;
+      }
+      stream.transition(Http2StreamState.Event.SEND_HEADERS);
+    }
+    if (responseWriter.dataBytesInBatch() != 0 || responseWriter.endStreamInBatch()) {
+      if (responseWriter.dataBytesInBatch() != 0) {
+        stream.transition(
+            responseWriter.endStreamInBatch() && !responseWriter.trailerHeadersInBatch()
+                ? Http2StreamState.Event.SEND_DATA_ES
+                : Http2StreamState.Event.SEND_DATA);
+      }
+      if (responseWriter.trailerHeadersInBatch()) {
+        stream.transition(Http2StreamState.Event.SEND_HEADERS_ES);
+      } else if (responseWriter.dataBytesInBatch() == 0 && responseWriter.endStreamInBatch()) {
+        stream.transition(Http2StreamState.Event.SEND_DATA_ES);
+      }
+    }
+  }
+
+  @Override
+  public void responseBatchCompleted(Http2Stream stream) {
+    stream.touch();
+    stream.endResponseBatch();
+    int streamId = stream.id();
+    if (streamId == 0) return;
+    if (stream.cancelled()) {
+      streams.remove(stream.id());
+      streams.release(stream);
+      return;
+    }
+    if (stream.responseWriter().finished()) {
+      if (stream.state() == Http2StreamState.CLOSED) {
+        if (!streams.retire(stream, streamId)) streams.release(stream);
+      }
+      return;
+    }
+    scheduleResume(stream);
+  }
+
+  private void detachFinalBatch(Http2Stream stream, Http2ResponseWriter writer) {
+    if (writer.finished() && stream.state() == Http2StreamState.CLOSED) {
+      streams.detach(stream, stream.id());
+    }
+  }
+
+  private void failAndRelease(Http2Stream stream, Http2ErrorCode error, Exception cause) {
+    int streamId = stream.id();
+    if (!streams.removeIfSame(stream, streamId) && stream.id() != streamId) return;
+    if (cause != null) log.error("HTTP/2 stream {} failed", streamId, cause);
+    try {
+      stream.cancel();
+    } catch (RuntimeException cancellationFailure) {
+      log.debug("Failed to cancel HTTP/2 stream {} cleanly", streamId, cancellationFailure);
+    }
+    try {
+      failures.fail(streamId, error);
+    } catch (IOException writeFailure) {
+      log.debug("Failed to write RST_STREAM for {}", streamId, writeFailure);
+    } finally {
+      streams.release(stream);
+    }
+  }
+}

flash/src/main/java/dev/relism/flash/http2/Http2StreamException.java

@@ -0,0 +1,43 @@
+package dev.relism.flash.http2;
+
+/**
+ * A <b>stream-level</b> HTTP/2 error, scoped to one stream id. Results in an {@code RST_STREAM}
+ * frame for {@link #streamId()} with {@link #errorCode()}; the connection and every other
+ * stream on it are unaffected. Compare {@link Http2Exception}, whose scope is the whole
+ * connection.
+ *
+ * <p>Deliberately does <b>not</b> extend {@link java.io.IOException}, for the same reason as
+ * {@link Http2Exception}: the connection loop must be able to distinguish "we decided to reject
+ * this stream" from "the socket failed" by catching unrelated exception types.
+ *
+ * <h3>Why this allocates, unlike {@code Http2Exception}'s singletons</h3>
+ * Every instance carries a distinct {@link #streamId()}, so it cannot be a shared singleton the
+ * exempts error paths. The scenario where this matters most — a peer opening and resetting
+ * thousands of streams per second (the Rapid Reset pattern, CVE-2023-44487) — is bounded by
+ * that can force RST_STREAM generation fast enough for GC pressure to matter has already
+ * tripped {@code Http2Limits.MAX_RESET_STREAMS_PER_INTERVAL} and the connection is being torn
+ * down anyway.
+ *
+ * <p>Stack trace capture is disabled for the same cost reason as {@link Http2Exception}.
+ */
+public final class Http2StreamException extends RuntimeException {
+
+    private final Http2ErrorCode errorCode;
+    private final int streamId;
+
+    public Http2StreamException(int streamId, Http2ErrorCode errorCode, String message) {
+        super(message, null, false, false);
+        this.streamId = streamId;
+        this.errorCode = errorCode;
+    }
+
+    /** The id of the stream this error terminates. */
+    public int streamId() {
+        return streamId;
+    }
+
+    /** The RFC 9113 §7 error code to send in the {@code RST_STREAM} frame. */
+    public Http2ErrorCode errorCode() {
+        return errorCode;
+    }
+}

flash/src/main/java/dev/relism/flash/http2/RollingWindowCounter.java

@@ -0,0 +1,31 @@
+package dev.relism.flash.http2;
+
+/** Allocation-free two-bucket rolling rate counter owned by one connection thread. */
+final class RollingWindowCounter {
+  private final long bucketNanos;
+  private long currentBucket;
+  private int currentCount;
+  private int previousCount;
+
+  RollingWindowCounter(long intervalMillis) {
+    if (intervalMillis < 2) throw new IllegalArgumentException("interval must be at least 2 ms");
+    bucketNanos = intervalMillis * 1_000_000L / 2;
+  }
+
+  boolean incrementExceeded(int limit) {
+    return incrementExceeded(limit, System.nanoTime());
+  }
+
+  boolean incrementExceeded(int limit, long nowNanos) {
+    long bucket = nowNanos / bucketNanos;
+    if (currentBucket == 0) {
+      currentBucket = bucket;
+    } else if (bucket != currentBucket) {
+      previousCount = bucket == currentBucket + 1 ? currentCount : 0;
+      currentCount = 0;
+      currentBucket = bucket;
+    }
+    currentCount++;
+    return currentCount + previousCount > limit;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/frame/FrameFlags.java

@@ -0,0 +1,39 @@
+package dev.relism.flash.http2.frame;
+
+/**
+ * The frame-header flag bits (RFC 9113 §6), as bitwise constants plus predicate helpers.
+ *
+ * <h3>The deliberate collision</h3>
+ * Bit {@code 0x1} means different things on different frame types: {@link #END_STREAM} on
+ * {@code DATA}/{@code HEADERS}, {@link #ACK} on {@code SETTINGS}/{@code PING}. They are the same
+ * bit position because the RFC defines flags per-type, not globally — reusing the numeric value
+ * is intentional on the wire, not a naming accident here. **Never call {@link #isEndStream} on a
+ * SETTINGS/PING frame's flags, or {@link #isAck} on a DATA/HEADERS frame's** — each predicate is
+ * named for the one frame type family it is valid to call it on; mixing them up silently
+ * misreads an unrelated bit rather than throwing, because the bit pattern is, by construction,
+ * identical.
+ *
+ * <p>RFC 9113 §4.1: flag bits not defined for a frame's type MUST be ignored on receipt and MUST
+ * NOT be set when sending. This class only ever tests bits it defines for the type the caller is
+ * working with; undefined bits are never inspected.
+ */
+public final class FrameFlags {
+    private FrameFlags() {}
+
+    /** DATA/HEADERS: no more frames will be sent for this stream in this direction. */
+    public static final int END_STREAM = 0x1;
+    /** SETTINGS/PING: this frame acknowledges the peer's own frame, rather than proposing new values. */
+    public static final int ACK = 0x1;
+    /** HEADERS/PUSH_PROMISE/CONTINUATION: the header block is complete — no CONTINUATION follows. */
+    public static final int END_HEADERS = 0x4;
+    /** DATA/HEADERS/PUSH_PROMISE: a pad-length byte and trailing padding are present — see {@link Padding}. */
+    public static final int PADDED = 0x8;
+    /** HEADERS: deprecated stream-dependency/weight fields are present (RFC 9113 §5.3.2 — parsed and discarded). */
+    public static final int PRIORITY = 0x20;
+
+    public static boolean isEndStream(int flags) { return (flags & END_STREAM) != 0; }
+    public static boolean isAck(int flags) { return (flags & ACK) != 0; }
+    public static boolean isEndHeaders(int flags) { return (flags & END_HEADERS) != 0; }
+    public static boolean isPadded(int flags) { return (flags & PADDED) != 0; }
+    public static boolean hasPriority(int flags) { return (flags & PRIORITY) != 0; }
+}

flash/src/main/java/dev/relism/flash/http2/frame/FrameHeader.java

@@ -0,0 +1,83 @@
+package dev.relism.flash.http2.frame;
+
+/**
+ * A <b>flyweight</b> over one frame's 9-byte header plus its payload location, both still living
+ * in {@link Http2FrameReader}'s own read buffer. One instance per connection, {@link #reset}
+ * in place by every {@link Http2FrameReader#readFrame()} call — never allocated per frame
+ * (mirrors the existing {@code WebSocketFrame} reuse idiom in {@code dev.relism.flash.websocket}).
+ *
+ * <h3>Lifetime contract</h3>
+ * Valid only until the next {@link Http2FrameReader#readFrame()}/{@code consumeFrame()} call on
+ * the same reader — same "do not retain past the handler" rule the rest of this codebase's
+ * buffer-backed flyweights (`Http1HeaderMap`, `WebSocketFrame`) already document. The payload bytes
+ * are also transient: whatever layer needs to retain a DATA frame's payload past this window
+ *
+ * <h3>Reserved bit and unknown types</h3>
+ * {@link #streamId()} has already had the wire's reserved high bit (RFC 9113 §4.1: "R: A
+ * reserved 1-bit field... The semantics of this bit are undefined, and the bit MUST be ignored
+ * when receiving") masked off during {@link #reset} — callers never see it and never need to
+ * mask it themselves. {@link #type()} is {@code null} for a type code {@link FrameType} does not
+ * recognise (i.e. {@code typeCode() > FrameType.maxKnown()}); per RFC 9113 §4.1 such frames must
+ * be ignored, not rejected — {@link #typeCode()} remains available so the caller can still log
+ * or count it before skipping the payload.
+ */
+public final class FrameHeader {
+    private byte[] buf;
+    private int length;
+    private int typeCode;
+    private FrameType type;
+    private int flags;
+    private int streamId;
+    private int payloadOffset;
+
+    /** Called by {@link Http2FrameReader} only, once the full 9-byte header is available at {@code buf[off]}. */
+    void reset(byte[] buf, int off) {
+        this.buf = buf;
+        int b0 = buf[off] & 0xFF, b1 = buf[off + 1] & 0xFF, b2 = buf[off + 2] & 0xFF;
+        this.length = (b0 << 16) | (b1 << 8) | b2;
+        this.typeCode = buf[off + 3] & 0xFF;
+        this.type = FrameType.fromCode(typeCode);
+        this.flags = buf[off + 4] & 0xFF;
+        // RFC 9113 §4.1: the top bit of byte 5 is reserved and MUST be ignored on receipt —
+        // masked here, once, rather than requiring every caller to remember to.
+        int b5 = buf[off + 5] & 0x7F;
+        int b6 = buf[off + 6] & 0xFF, b7 = buf[off + 7] & 0xFF, b8 = buf[off + 8] & 0xFF;
+        this.streamId = (b5 << 24) | (b6 << 16) | (b7 << 8) | b8;
+        this.payloadOffset = off + 9;
+    }
+
+    /** Payload length in bytes, as declared by the frame header (0..2^24-1 before any limit check). */
+    public int length() {
+        return length;
+    }
+
+    /** The raw wire type byte, valid even when {@link #type()} is {@code null} (an unrecognised type). */
+    public int typeCode() {
+        return typeCode;
+    }
+
+    /** The recognised frame type, or {@code null} if {@link #typeCode()} is not one of RFC 9113's 10. */
+    public FrameType type() {
+        return type;
+    }
+
+    /** The raw flags byte — interpret via {@link FrameFlags}, which is type-specific. */
+    public int flags() {
+        return flags;
+    }
+
+    /** Stream identifier, reserved bit already masked. {@code 0} means "the connection itself". */
+    public int streamId() {
+        return streamId;
+    }
+
+    /** The backing buffer — see the class Javadoc's lifetime contract before retaining a reference. */
+    public byte[] buffer() {
+        return buf;
+    }
+
+    /** Offset of the first payload byte within {@link #buffer()}. Payload spans {@code [payloadOffset(), payloadOffset() + length())}. */
+    public int payloadOffset() {
+        return payloadOffset;
+    }
+}

flash/src/main/java/dev/relism/flash/http2/frame/FrameType.java

@@ -0,0 +1,113 @@
+package dev.relism.flash.http2.frame;
+
+/**
+ * The 10 HTTP/2 frame types (RFC 9113 §6), plus the shared per-type validation rules {@link
+ * FrameValidator} enforces. Values above {@code 0x9} are not assigned a constant here — RFC 9113
+ * §4.1 requires unknown types to be silently ignored (read and discard the payload), which {@link
+ * Http2FrameReader}'s caller implements by checking {@code type > FrameType.maxKnown()} rather than
+ * by this enum growing an {@code UNKNOWN} member (an {@code UNKNOWN} constant would misleadingly
+ * suggest "a recognised category of unrecognised frame", when the correct handling is simply "not
+ * this table, skip it").
+ *
+ * <p>Each constant carries the RFC-mandated payload length bounds, whether a zero stream id is
+ * required/forbidden/either, and whether the frame counts toward the CONTINUATION-flood guard
+ * (bounded by {@code Http2Limits#MAX_CONTINUATION_FRAMES_PER_BLOCK}) — see {@link FrameValidator}
+ * for how these are applied and the specific RFC citation per rule.
+ */
+public enum FrameType {
+  /** RFC 9113 §6.1. Stream body bytes. Stream id required. Length: 0..MAX_FRAME_SIZE. */
+  DATA(0x0, 0, Integer.MAX_VALUE, StreamIdRule.REQUIRED),
+  /** RFC 9113 §6.2. Header block fragment (HPACK). Stream id required. */
+  HEADERS(0x1, 0, Integer.MAX_VALUE, StreamIdRule.REQUIRED),
+  PRIORITY(0x2, 5, 5, StreamIdRule.REQUIRED),
+  /** RFC 9113 §6.4. Stream-level error. Exactly 4 bytes (the error code). Stream id required. */
+  RST_STREAM(0x3, 4, 4, StreamIdRule.REQUIRED),
+  /**
+   * RFC 9113 §6.5. Connection-level parameters. Length must be a multiple of 6. Stream id must be
+   * 0.
+   */
+  SETTINGS(0x4, 0, Integer.MAX_VALUE, StreamIdRule.FORBIDDEN),
+  /**
+   * RFC 9113 §6.6. Never sent (Flash advertises {@code SETTINGS_ENABLE_PUSH=0}); receiving one from
+   * a client is a protocol error.
+   */
+  PUSH_PROMISE(0x5, 4, Integer.MAX_VALUE, StreamIdRule.REQUIRED),
+  /**
+   * RFC 9113 §6.7. Connection liveness / RTT probe. Exactly 8 bytes of opaque data. Stream id must
+   * be 0.
+   */
+  PING(0x6, 8, 8, StreamIdRule.FORBIDDEN),
+  /**
+   * RFC 9113 §6.8. Connection shutdown notice. At least 8 bytes (last-stream-id + error code).
+   * Stream id must be 0.
+   */
+  GOAWAY(0x7, 8, Integer.MAX_VALUE, StreamIdRule.FORBIDDEN),
+  /**
+   * RFC 9113 §6.9. Flow-control window increment. Exactly 4 bytes. Stream id may be either (0 =
+   * connection window).
+   */
+  WINDOW_UPDATE(0x8, 4, 4, StreamIdRule.EITHER),
+  /**
+   * RFC 9113 §6.10. Continuation of a header block that did not fit one HEADERS/PUSH_PROMISE frame.
+   * Stream id required.
+   */
+  CONTINUATION(0x9, 0, Integer.MAX_VALUE, StreamIdRule.REQUIRED);
+
+  /** Whether a frame type requires stream id 0, requires it non-zero, or permits either. */
+  public enum StreamIdRule {
+    REQUIRED,
+    FORBIDDEN,
+    EITHER
+  }
+
+  private static final FrameType[] BY_CODE = new FrameType[values().length];
+
+  static {
+    for (FrameType t : values()) {
+      BY_CODE[t.code] = t;
+    }
+  }
+
+  private final int code;
+  private final int minLength;
+  private final int maxLength;
+  private final StreamIdRule streamIdRule;
+
+  FrameType(int code, int minLength, int maxLength, StreamIdRule streamIdRule) {
+    this.code = code;
+    this.minLength = minLength;
+    this.maxLength = maxLength;
+    this.streamIdRule = streamIdRule;
+  }
+
+  public int code() {
+    return code;
+  }
+
+  public int minLength() {
+    return minLength;
+  }
+
+  public int maxLength() {
+    return maxLength;
+  }
+
+  public StreamIdRule streamIdRule() {
+    return streamIdRule;
+  }
+
+  /**
+   * The highest type code this enum recognises — anything above must be ignored per RFC 9113 §4.1.
+   */
+  public static int maxKnown() {
+    return CONTINUATION.code;
+  }
+
+  /**
+   * Looks up the constant for a wire type byte, or {@code null} if it is an unrecognised
+   * (to-be-ignored) type.
+   */
+  public static FrameType fromCode(int code) {
+    return (code >= 0 && code < BY_CODE.length) ? BY_CODE[code] : null;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/frame/FrameValidator.java

@@ -0,0 +1,88 @@
+package dev.relism.flash.http2.frame;
+
+import dev.relism.flash.http2.Http2ErrorCode;
+import dev.relism.flash.http2.Http2Exception;
+import dev.relism.flash.http2.Http2Limits;
+
+/**
+ * Table-driven RFC 9113 per-frame-type validation: length bounds, the stream-id
+ * required/forbidden/either rule, and the two special-cased structural rules ({@code SETTINGS}'
+ * multiple-of-6 length, {@code PUSH_PROMISE} always rejected from a client) that do not fit a
+ * class is the code that reads it.
+ *
+ * <p><b>The error code is not uniform</b> — read the RFC per violation, not just per type. A
+ * {@code SETTINGS} frame with a bad length is {@code FRAME_SIZE_ERROR}; the same frame with a
+ * non-zero stream id is {@code PROTOCOL_ERROR}. This class throws the specific code each
+ * violation's own RFC citation requires, not a single blanket code per type.
+ */
+public final class FrameValidator {
+    private FrameValidator() {}
+
+    /**
+     * Validates {@code header} against RFC 9113's rules for its type.
+     *
+     * @param insideHeaderBlock whether this frame arrived between a HEADERS/PUSH_PROMISE frame
+     *                          lacking {@code END_HEADERS} and its terminating CONTINUATION —
+     *                          changes the handling of an unrecognised type (§6.10: a
+     *                          {@code PROTOCOL_ERROR}, not the usual silent ignore, since an
+     *                          in-progress header block cannot tolerate an interloper frame of
+     *                          any kind without desynchronizing HPACK's stateful decode)
+     * @throws Http2Exception on any RFC violation, with the specific error code the violated
+     *         rule mandates
+     */
+    public static void validate(FrameHeader header, boolean insideHeaderBlock) {
+        FrameType type = header.type();
+
+        if (type == null) {
+            // RFC 9113 §4.1: unknown frame types MUST be ignored — except inside an in-progress
+            // header block (§6.10), where anything other than CONTINUATION desynchronizes HPACK.
+            if (insideHeaderBlock) {
+                throw Http2Exception.of(Http2ErrorCode.PROTOCOL_ERROR,
+                        "unrecognised frame type " + header.typeCode() + " received inside a header block");
+            }
+            return;
+        }
+
+        int length = header.length();
+
+        // RFC 9113 §6.5: a SETTINGS frame's length MUST be a multiple of 6 (each entry is a
+        // 2-byte identifier + 4-byte value). Checked before the generic bounds below, since the
+        // generic table only expresses a min/max range, not a modulus.
+        if (type == FrameType.SETTINGS && length % 6 != 0) {
+            throw Http2Exception.FRAME_SIZE_ERROR;
+        }
+
+        if (length < type.minLength() || length > type.maxLength()) {
+            throw Http2Exception.FRAME_SIZE_ERROR;
+        }
+
+        // Redundant with Http2FrameReader's own pre-allocation check for frames it read itself,
+        // but this method must also be correct for a FrameHeader built any other way (tests,
+        // and — in later phases — frames reassembled from multiple reads), so the bound is
+        // re-asserted here rather than trusted from the caller.
+        if (length > Http2Limits.MAX_FRAME_SIZE_LOCAL) {
+            throw Http2Exception.FRAME_SIZE_ERROR;
+        }
+
+        int streamId = header.streamId();
+        switch (type.streamIdRule()) {
+            case REQUIRED -> {
+                if (streamId == 0) {
+                    throw Http2Exception.of(Http2ErrorCode.PROTOCOL_ERROR, type + " requires a non-zero stream id");
+                }
+            }
+            case FORBIDDEN -> {
+                if (streamId != 0) {
+                    throw Http2Exception.of(Http2ErrorCode.PROTOCOL_ERROR, type + " must have stream id 0, got " + streamId);
+                }
+            }
+            case EITHER -> { /* WINDOW_UPDATE: 0 (connection window) or non-zero (stream window) both valid */ }
+        }
+
+        // (Flash advertises SETTINGS_ENABLE_PUSH=0 and never sends one); receiving one at all
+        // means the peer believes it is talking to a client, which is always a protocol error.
+        if (type == FrameType.PUSH_PROMISE) {
+            throw Http2Exception.of(Http2ErrorCode.PROTOCOL_ERROR, "PUSH_PROMISE received from a client");
+        }
+    }
+}

flash/src/main/java/dev/relism/flash/http2/frame/FrameWriteBuffer.java

@@ -0,0 +1,75 @@
+package dev.relism.flash.http2.frame;
+
+import dev.relism.flash.bytes.ByteWriter;
+
+/**
+ * Serializes HTTP/2 frames into a {@link ByteWriter} scratch buffer with the standard
+ * length-back-patching technique: {@link #beginFrame} writes a 9-byte header with a placeholder
+ * length, the caller writes the payload directly through {@link #writer()} (the same
+ * {@link ByteWriter}), and {@link #endFrame} rewrites the length once it is known — the payload
+ * size is rarely known before it is serialized (an HPACK-encoded header block, in particular,
+ * has no cheap way to be measured in advance).
+ *
+ * issues one bulk {@code write}, rather than streaming bytes as they are produced: streaming
+ * would require knowing the length <em>before</em> the first byte goes out, which back-patching
+ * deliberately avoids needing.
+ *
+ * <h3>Usage</h3>
+ * <pre>{@code
+ * FrameWriteBuffer out = new FrameWriteBuffer(new ByteWriter(4096));
+ * out.beginFrame(FrameType.SETTINGS, 0, 0);
+ * out.writer().writeUInt16(SETTINGS_MAX_CONCURRENT_STREAMS);
+ * out.writer().writeUInt32(100);
+ * out.endFrame();
+ * // out.writer().array()[0, out.writer().length()) now holds one complete, correctly-lengthed frame
+ * }</pre>
+ *
+ * <h3>Multiple frames, one buffer</h3>
+ * {@link #beginFrame}/{@link #endFrame} pairs may be repeated on the same instance without a
+ * {@link ByteWriter#reset()} between them — each pair appends one more complete frame after
+ * whatever was already written, which is exactly what {@link Http2FrameWriter#write} wants for a
+ * single bulk write covering several frames (e.g. HEADERS followed immediately by its first
+ * DATA frame).
+ *
+ * <h3>Thread-safety</h3>
+ * Not thread-safe — exactly one writer at a time, the same convention every other per-connection
+ * scratch object in this codebase follows.
+ */
+public final class FrameWriteBuffer {
+    private final ByteWriter writer;
+    private int headerStart = -1;
+
+    public FrameWriteBuffer(ByteWriter writer) {
+        this.writer = writer;
+    }
+
+    /** The underlying {@link ByteWriter} — write the frame's payload directly through this between {@link #beginFrame} and {@link #endFrame}. */
+    public ByteWriter writer() {
+        return writer;
+    }
+
+    /** Writes a 9-byte frame header with a placeholder length, to be filled in by {@link #endFrame}. */
+    public void beginFrame(FrameType type, int flags, int streamId) {
+        if (headerStart != -1) {
+            throw new IllegalStateException("beginFrame() called again before the previous frame's endFrame()");
+        }
+        headerStart = writer.length();
+        writer.writeUInt24(0); // length placeholder
+        writer.writeByte((byte) type.code());
+        writer.writeByte((byte) flags);
+        writer.writeUInt31(streamId);
+    }
+
+    /** Back-patches the length field written by {@link #beginFrame} now that the payload's size is known. */
+    public void endFrame() {
+        if (headerStart == -1) {
+            throw new IllegalStateException("endFrame() called without a matching beginFrame()");
+        }
+        int payloadLength = writer.length() - (headerStart + 9);
+        byte[] buf = writer.array();
+        buf[headerStart] = (byte) (payloadLength >>> 16);
+        buf[headerStart + 1] = (byte) (payloadLength >>> 8);
+        buf[headerStart + 2] = (byte) payloadLength;
+        headerStart = -1;
+    }
+}

flash/src/main/java/dev/relism/flash/http2/frame/Http2FrameReader.java

@@ -0,0 +1,160 @@
+package dev.relism.flash.http2.frame;
+
+import dev.relism.flash.http2.Http2Exception;
+import dev.relism.flash.http2.Http2Limits;
+import dev.relism.flash.transport.BufferedByteSource;
+import java.io.EOFException;
+import java.io.IOException;
+import java.util.Arrays;
+
+/**
+ * Reads length-prefixed HTTP/2 frames from one connection's {@link BufferedByteSource}. Simpler
+ * than {@code RequestParser} by construction: HTTP/2 frames declare their length up front (the
+ * 9-byte header), so nothing is ever scanned for — {@code Http2FrameReader} only ever needs to know
+ * "do I have N bytes yet", never "where does this end".
+ *
+ * <h3>Buffer discipline</h3>
+ *
+ * One growable {@code byte[]} per connection, reused across every frame — the same
+ * compact-before-grow discipline {@code RequestParser}'s own buffer uses. A frame's declared length
+ * is checked against {@link Http2Limits#MAX_FRAME_SIZE_LOCAL} <em>before</em> the buffer
+ * length-check, not after an allocation already paid for it.
+ *
+ * <h3>Usage</h3>
+ *
+ * <pre>{@code
+ * FrameHeader header = reader.readFrame();
+ * if (header == null) { /* clean EOF between frames — connection closing *\/ }
+ * // ... process header.buffer()[header.payloadOffset(), +header.length()) ...
+ * reader.consumeFrame(); // MUST be called before the next readFrame()
+ * }</pre>
+ *
+ * <h3>Thread-safety</h3>
+ *
+ * Not thread-safe — exactly one virtual thread (the connection's demux loop) ever calls this, the
+ * same invariant every other per-connection reader in this codebase assumes.
+ */
+public final class Http2FrameReader {
+  private static final int FRAME_HEADER_SIZE = 9;
+  private static final int INITIAL_BUFFER_SIZE = 16 * 1024;
+
+  private final BufferedByteSource in;
+  private final FrameHeader header = new FrameHeader();
+  private byte[] buffer;
+  private int base; // offset of the first unconsumed byte
+  private int totalRead; // count of valid unconsumed bytes at [base, base + totalRead)
+  private long frameDeadlineNanos;
+
+  public Http2FrameReader(BufferedByteSource in) {
+    this(in, INITIAL_BUFFER_SIZE);
+  }
+
+  public Http2FrameReader(BufferedByteSource in, int initialBufferSize) {
+    this.in = in;
+    this.buffer = new byte[Math.max(initialBufferSize, FRAME_HEADER_SIZE)];
+  }
+
+  /**
+   * Reads the next frame's header and payload, bounded by {@link
+   * Http2Limits#FRAME_READ_TIMEOUT_MS}, and returns the reused {@link FrameHeader} flyweight
+   * positioned over it — or {@code null} on a clean EOF between frames (the peer closed the
+   * connection while nothing was in flight; not an error).
+   *
+   * <p>The caller MUST call {@link #consumeFrame()} exactly once after processing this frame (or
+   * deciding to discard it) and before calling this method again.
+   *
+   * @throws Http2Exception if the declared length exceeds {@link Http2Limits#MAX_FRAME_SIZE_LOCAL}
+   * @throws EOFException if the connection closes after a frame has already started arriving
+   * @throws java.net.SocketTimeoutException if {@link Http2Limits#FRAME_READ_TIMEOUT_MS} elapses
+   */
+  public FrameHeader readFrame() throws IOException {
+    return readFrame(Http2Limits.FRAME_READ_TIMEOUT_MS);
+  }
+
+  /** Reads one frame using a caller-supplied upper bound for this frame's absolute deadline. */
+  public FrameHeader readFrame(long timeoutMs) throws IOException {
+    if (timeoutMs <= 0) throw new IllegalArgumentException("timeoutMs must be positive");
+    long now = System.nanoTime();
+    if (frameDeadlineNanos == 0) {
+      frameDeadlineNanos = now + Http2Limits.FRAME_READ_TIMEOUT_MS * 1_000_000L;
+    }
+    in.setDeadline(Math.min(frameDeadlineNanos, now + timeoutMs * 1_000_000L));
+    try {
+      if (!ensureAvailable(FRAME_HEADER_SIZE)) {
+        frameDeadlineNanos = 0;
+        return null; // clean EOF: nothing buffered yet, peer closed between frames
+      }
+      int declaredLength = decodeLength(buffer, base);
+      // never causes an oversized allocation, only a rejection.
+      if (declaredLength > Http2Limits.MAX_FRAME_SIZE_LOCAL) {
+        throw Http2Exception.FRAME_SIZE_ERROR;
+      }
+      ensureAvailable(FRAME_HEADER_SIZE + declaredLength);
+      header.reset(buffer, base);
+      return header;
+    } catch (java.net.SocketTimeoutException timeout) {
+      if (totalRead == 0) frameDeadlineNanos = 0;
+      throw timeout;
+    } finally {
+      in.clearDeadline();
+    }
+  }
+
+  /** Advances past the frame last returned by {@link #readFrame()}. Zero-copy, zero-allocation. */
+  public void consumeFrame() {
+    int consumed = FRAME_HEADER_SIZE + header.length();
+    base += consumed;
+    totalRead -= consumed;
+    if (totalRead == 0) {
+      base = 0; // nothing buffered — reset to the front rather than drifting forever
+    }
+    frameDeadlineNanos = 0;
+  }
+
+  /** Whether a partially received frame exhausted its non-renewable absolute deadline. */
+  public boolean frameDeadlineExpired() {
+    return totalRead != 0 && System.nanoTime() >= frameDeadlineNanos;
+  }
+
+  /** Whether another frame may be consumed immediately without waiting for network input. */
+  public boolean hasBufferedInput() {
+    return totalRead != 0 || in.available() != 0;
+  }
+
+  private static int decodeLength(byte[] buf, int off) {
+    int b0 = buf[off] & 0xFF, b1 = buf[off + 1] & 0xFF, b2 = buf[off + 2] & 0xFF;
+    return (b0 << 16) | (b1 << 8) | b2;
+  }
+
+  /**
+   * Ensures at least {@code need} bytes are available starting at {@link #base}, growing or
+   * compacting the buffer as necessary. Returns {@code false} only for a clean EOF with nothing at
+   * all buffered yet (the between-frames case); an EOF after any bytes of the current frame have
+   * already arrived is a genuine truncation and throws.
+   */
+  private boolean ensureAvailable(int need) throws IOException {
+    while (totalRead < need) {
+      if (base + need > buffer.length) {
+        if (base > 0) {
+          // Compact: slide unconsumed bytes to the front — frees room without growing.
+          System.arraycopy(buffer, base, buffer, 0, totalRead);
+          base = 0;
+        } else {
+          // need <= 9 + MAX_FRAME_SIZE_LOCAL always, by readFrame()'s own check before
+          // the payload-sized call — grow exactly enough, never unbounded.
+          int grown = buffer.length;
+          while (grown < need) grown *= 2;
+          buffer = Arrays.copyOf(buffer, grown);
+        }
+      }
+      int n = in.read(buffer, base + totalRead, buffer.length - base - totalRead);
+      if (n < 0) {
+        if (totalRead == 0) return false;
+        throw new EOFException(
+            "connection closed mid-frame (" + totalRead + "/" + need + " bytes read)");
+      }
+      totalRead += n;
+    }
+    return true;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/frame/Http2FrameWriter.java

@@ -0,0 +1,283 @@
+package dev.relism.flash.http2.frame;
+
+import dev.relism.flash.http2.Http2Limits;
+import java.io.IOException;
+import java.io.InterruptedIOException;
+import java.util.Set;
+import java.util.concurrent.ConcurrentHashMap;
+import java.util.concurrent.locks.ReentrantLock;
+
+/**
+ * The one component every HTTP/2 write in this codebase passes through — connection frames and
+ * stream frames alike (both are just {@link WriteIntent}s). Its entire job is serializing
+ * concurrent access to one connection's socket write side as cheaply as physically possible,
+ * because under multiplexing every stream on a connection shares that one socket.
+ *
+ * <h2>The design, three layers</h2>
+ *
+ * <p><b>Layer 1 — serialize outside the lock.</b> By the time {@link #write} is called, the caller
+ * has already built its complete frame into a buffer it owns (see {@link WriteIntent}). This writer
+ * never serializes anything; it only ever issues one bulk {@code sink.write(buffer, offset,
+ * length)} call while holding the lock — never many small writes, which would turn "hold the lock"
+ * into "hold the lock across a serialization pass."
+ *
+ * <p><b>Layer 2 — {@link ReentrantLock}, never {@code synchronized}.</b> On Java 21, a virtual
+ * thread blocking inside {@code synchronized} pins its carrier platform thread; blocking on a
+ * {@link ReentrantLock} unmounts it instead (JEP 491, which removes the {@code synchronized} {@code
+ * ReentrantLock} is also load-bearing here for a second reason {@code synchronized} cannot offer:
+ * {@link ReentrantLock#tryLock()}.
+ *
+ * <p><b>Layer 3 — {@code tryLock()} fast path, intrusive MPSC fallback.</b> The overwhelmingly
+ * common case, even on a genuinely multiplexed connection, is exactly one stream wanting to write
+ * at a given instant. {@code tryLock()} on an uncontended lock is one successful CAS; the calling
+ * thread writes inline and releases — no handoff, no queue touched, no allocation, no context
+ * switch. Only when {@code tryLock()} fails (genuine contention) does the intent get published
+ * through {@link IntrusiveMpscQueue} (one more CAS, still zero allocation — the intent itself is
+ * the queue node) for the current lock holder to drain.
+ *
+ * <h3>Lost-wakeup avoidance</h3>
+ *
+ * The classic hazard: a producer offers its intent to the queue at the exact moment the current
+ * holder has just found the queue empty and is about to unlock — the item would be stranded with
+ * nobody left to drain it. This is closed by two cooperating checks, and the correctness argument
+ * for why together they are sufficient is a happens-before chain through the queue's {@code
+ * AtomicReference} and the lock's own acquire/release ordering (recorded in full in {@code
+ * WRITER.md}, since it is exactly the kind of reasoning a future reader must be able to re-derive,
+ * not just trust):
+ *
+ * <pre>
+ * write(intent):
+ *   if tryLock() succeeds:           // 1 CAS, the fast path
+ *       drive(intent)                // write intent directly, then drain the queue, then unlock
+ *   else:
+ *       queue.offer(intent)          // 1 CAS, zero allocation
+ *       if tryLock() succeeds:       // the producer's own second chance
+ *           drive(null)              // drain whatever is queued, including our own intent
+ *
+ * drive(firstIntentOrNull):
+ *   write firstIntentOrNull if present, then poll-and-write until the queue is empty
+ *   unlock()
+ *   while queue.hasWork():           // the re-check-after-unlock that closes the race
+ *       if !tryLock(): break         // someone else is now responsible; their own recheck covers us
+ *       poll-and-write until empty
+ *       unlock()
+ * </pre>
+ *
+ * A frame's bytes are never interleaved with another frame's bytes: every write of one intent is a
+ * single {@code sink.write} call issued while holding the lock, and the lock is not released
+ * between a {@code WriteIntent}'s bytes.
+ *
+ * <h3>Write timeout</h3>
+ *
+ * A blocking write is unavoidable when the kernel send buffer is full and the peer is not reading —
+ * whoever holds the lock is blocked in the syscall, holding up every other stream on the
+ * connection. This is bounded by {@link Http2Limits#WRITE_TIMEOUT_MS}, enforced by a shared
+ * background reaper ({@link WriteTimeoutReaper}) that interrupts the blocked thread past the
+ * deadline — {@code Socket#setSoTimeout} bounds reads, not writes, so it cannot be used here.
+ * connection setup), so arming/disarming the deadline for each individual write is two {@code
+ * volatile} field writes, not an allocation.
+ */
+public final class Http2FrameWriter {
+
+  /**
+   * What a frame's serialized bytes are ultimately written to. Kept minimal and separate from
+   * {@code java.io.OutputStream} so this class is testable without a real socket.
+   */
+  public interface Sink {
+    void write(byte[] buf, int off, int len) throws IOException;
+  }
+
+  private final Sink sink;
+  private final long writeTimeoutMs;
+  private final ReentrantLock lock = new ReentrantLock();
+  private final IntrusiveMpscQueue priorityQueue = new IntrusiveMpscQueue();
+  private final IntrusiveMpscQueue queue = new IntrusiveMpscQueue();
+
+  // Set only for the duration of an in-flight sink.write() call; see WriteTimeoutReaper. A
+  // single volatile write to arm, one to disarm — no timestamp is recorded here (see the
+  // reaper's own Javadoc for why: a per-write System.nanoTime() call measurably missed the
+  // N=1 gate's 50 ns overhead budget when this was first benchmarked, recorded in WRITER.md).
+  private volatile Thread writingThread;
+
+  public Http2FrameWriter(Sink sink) {
+    this(sink, Http2Limits.WRITE_TIMEOUT_MS);
+  }
+
+  public Http2FrameWriter(Sink sink, long writeTimeoutMs) {
+    this.sink = sink;
+    this.writeTimeoutMs = writeTimeoutMs;
+    WriteTimeoutReaper.register(this);
+  }
+
+  /**
+   * Serializes and writes one frame. Returns when the bytes are in the socket buffer or safely
+   * queued behind another writer. Never blocks on another stream's I/O while holding the lock for
+   * longer than that stream's own single bulk write.
+   *
+   * <p><b>Why the fast path is gated on {@code !queue.hasWork()}, not just {@code tryLock()}</b>
+   * Writing {@code intent} immediately, before anything already queued, is only safe when nothing
+   * is already queued. Without the {@code hasWork()} check, this sequence is possible — and
+   * violates same-producer ordering, which the stress test asserts: a producer's {@code write(a)}
+   * then {@code write(b)} contends and both get queued (fire-and-forget); the current holder is
+   * about to drain them but has not yet; that producer's very next call, {@code write(c)}, finds
+   * the lock free (the holder released it between the producer's calls) and would otherwise write
+   * {@code c} directly — landing on the wire before {@code a} and {@code b}, which are still
+   * sitting in the queue. Checking {@code hasWork()} first means "bypass the queue" only happens
+   * when the queue is observed genuinely empty, i.e. everything previously offered — by any
+   * producer — has already been written; see {@code WRITER.md} for the full argument.
+   */
+  public void write(WriteIntent intent) throws IOException {
+    if (!priorityQueue.hasWork() && !queue.hasWork() && lock.tryLock()) {
+      drive(intent);
+    } else {
+      queue.offer(intent);
+      if (lock.tryLock()) {
+        drive(null);
+      }
+    }
+  }
+
+  /**
+   * Writes a connection-control frame ahead of queued stream data. An already executing socket
+   * write is never interrupted, but once it completes the priority queue is drained before the
+   * ordinary queue. This is used for PING acknowledgements, SETTINGS acknowledgements, GOAWAY and
+   * RST_STREAM.
+   */
+  public void writePriority(WriteIntent intent) throws IOException {
+    priorityQueue.offer(intent);
+    if (lock.tryLock()) {
+      drive(null);
+    }
+  }
+
+  /**
+   * Flushes any queued intents. Called by the demux loop when it has nothing left to read — a no-op
+   * on the (overwhelmingly common) fast path where nothing is queued.
+   */
+  public void drain() throws IOException {
+    if (!priorityQueue.hasWork() && !queue.hasWork()) return;
+    if (lock.tryLock()) {
+      drive(null);
+    }
+  }
+
+  /**
+   * Deregisters this writer from the write-timeout reaper. Call once, when the connection closes.
+   */
+  public void close() {
+    WriteTimeoutReaper.unregister(this);
+  }
+
+  private void drive(WriteIntent firstIntentOrNull) throws IOException {
+    try {
+      if (firstIntentOrNull != null) writeDirect(firstIntentOrNull);
+      drainQueues();
+    } finally {
+      lock.unlock();
+    }
+    // Lost-wakeup fix: re-check after unlocking, looping because this cycle itself can race
+    // the same way — see the class Javadoc for the correctness argument.
+    while (priorityQueue.hasWork() || queue.hasWork()) {
+      if (!lock.tryLock()) break;
+      try {
+        drainQueues();
+      } finally {
+        lock.unlock();
+      }
+    }
+  }
+
+  private void drainQueues() throws IOException {
+    WriteIntent next;
+    while (true) {
+      while ((next = priorityQueue.poll()) != null) {
+        writeDirect(next);
+      }
+      next = queue.poll();
+      if (next == null) return;
+      writeDirect(next);
+    }
+  }
+
+  private void writeDirect(WriteIntent intent) throws IOException {
+    writingThread = Thread.currentThread();
+    try {
+      sink.write(intent.buffer(), intent.offset(), intent.length());
+    } catch (IOException e) {
+      if (Thread.interrupted()) {
+        InterruptedIOException timeout =
+            new InterruptedIOException("HTTP/2 write timed out after ~" + writeTimeoutMs + " ms");
+        timeout.initCause(e);
+        throw timeout;
+      }
+      throw e;
+    } finally {
+      writingThread = null;
+      Thread.interrupted(); // clear a stray interrupt flag defensively before returning control
+      intent.completed();
+    }
+  }
+
+  /**
+   * A single shared daemon thread scanning every registered {@link Http2FrameWriter} for a blocking
+   * write that has overrun {@link Http2Limits#WRITE_TIMEOUT_MS}. One thread for the whole process
+   * (like {@code DateHeader}'s refresher), not one per connection — registration per-write one.
+   *
+   * <p>Deliberately does <em>not</em> ask each write to record a {@code System.nanoTime()} {@code
+   * nanoTime()} call (plus the extra volatile field it required) costing enough to miss the N=1
+   * gate's 50 ns-over-baseline budget (recorded in {@code WRITER.md}). Instead, the reaper counts
+   * <em>consecutive scans</em> a given writer has been observed still blocked ({@link
+   * #writingThread} non-null); a writer blocked for more than {@code WRITE_TIMEOUT_MS /
+   * SCAN_INTERVAL_MS} consecutive scans is interrupted. This trades a little precision (up to one
+   * scan interval of slop — already inherent to any background-reaper design) for removing all
+   * per-write timing cost.
+   */
+  static final class WriteTimeoutReaper {
+    private static final long SCAN_INTERVAL_MS = 50;
+    private static final Set<Http2FrameWriter> ACTIVE = ConcurrentHashMap.newKeySet();
+    // Touched only by the single reaper thread -- no synchronization needed.
+    private static final java.util.Map<Http2FrameWriter, Integer> BLOCKED_SCAN_COUNTS =
+        new java.util.IdentityHashMap<>();
+
+    static {
+      Thread reaper =
+          new Thread(
+              () -> {
+                while (true) {
+                  try {
+                    Thread.sleep(SCAN_INTERVAL_MS);
+                  } catch (InterruptedException e) {
+                    Thread.currentThread().interrupt();
+                    return;
+                  }
+                  for (Http2FrameWriter writer : ACTIVE) {
+                    Thread t = writer.writingThread;
+                    if (t == null) {
+                      BLOCKED_SCAN_COUNTS.remove(writer);
+                      continue;
+                    }
+                    int scans = BLOCKED_SCAN_COUNTS.merge(writer, 1, Integer::sum);
+                    long thresholdScans = Math.max(1, writer.writeTimeoutMs / SCAN_INTERVAL_MS);
+                    if (scans >= thresholdScans) {
+                      t.interrupt();
+                      BLOCKED_SCAN_COUNTS.remove(writer);
+                    }
+                  }
+                }
+              },
+              "flash-http2-write-timeout-reaper");
+      reaper.setDaemon(true);
+      reaper.start();
+    }
+
+    private WriteTimeoutReaper() {}
+
+    static void register(Http2FrameWriter writer) {
+      ACTIVE.add(writer);
+    }
+
+    static void unregister(Http2FrameWriter writer) {
+      ACTIVE.remove(writer);
+    }
+  }
+}

flash/src/main/java/dev/relism/flash/http2/frame/IntrusiveMpscQueue.java

@@ -0,0 +1,101 @@
+package dev.relism.flash.http2.frame;
+
+import java.util.concurrent.atomic.AtomicReference;
+
+/**
+ * A Vyukov-style intrusive multi-producer, single-consumer queue of {@link WriteIntent}s.
+ * "Intrusive" means the queued object <em>is</em> the node — {@link WriteIntent#mpscNext()} /
+ * {@link WriteIntent#setMpscNext} supply the linkage — so {@link #offer} allocates nothing: one
+ * {@link AtomicReference#getAndSet} CAS and that is the entire cost.
+ *
+ * <h3>Only {@link Http2FrameWriter} calls {@link #poll()}</h3>
+ * This queue is safe for any number of concurrent {@link #offer} callers, but {@link #poll()}
+ * must only ever be called by the single thread currently holding the writer's lock — exactly
+ * the invariant {@code Http2FrameWriter} maintains (it never calls {@code poll()} without
+ * holding the lock). Calling {@code poll()} from two threads concurrently is undefined.
+ *
+ * <h3>The stub node and the "inconsistent" result</h3>
+ * The queue always contains at least one node — a private, singleton {@code stub} — which lets
+ * {@link #offer} and {@link #poll} both proceed without ever observing a literal {@code null}
+ * head. A subtlety of this algorithm (documented here because it surprises readers unfamiliar
+ * with it, and it is the reason {@code Http2FrameWriter}'s drain loop is itself a loop, not a
+ * single pass): {@link #poll()} can return {@code null} even when {@link #offer} has completed
+ * and is "logically" enqueued, if that producer's {@code getAndSet} (which publishes the new
+ * tail pointer) has completed but its following {@code setMpscNext} (which links the *previous*
+ * tail to it) has not yet landed. This is a momentary, self-correcting race — the next
+ * {@code poll()} call (even from the same thread, immediately after) will see it — never a
+ * permanent loss. {@code Http2FrameWriter}'s lost-wakeup-avoidance protocol (see its Javadoc)
+ * already retries in exactly the way this requires.
+ */
+final class IntrusiveMpscQueue {
+
+    /**
+     * Sentinel node that is never returned by {@link #poll()} and never appears anywhere except
+     * internally. Its own {@code mpscNext} field is the only piece of mutable state on it.
+     */
+    private static final class Stub implements WriteIntent {
+        private volatile WriteIntent next;
+
+        @Override public byte[] buffer() { throw new UnsupportedOperationException("stub node"); }
+        @Override public int offset() { throw new UnsupportedOperationException("stub node"); }
+        @Override public int length() { throw new UnsupportedOperationException("stub node"); }
+        @Override public WriteIntent mpscNext() { return next; }
+        @Override public void setMpscNext(WriteIntent next) { this.next = next; }
+    }
+
+    private final Stub stub = new Stub();
+    private final AtomicReference<WriteIntent> head = new AtomicReference<>(stub);
+    private WriteIntent tail = stub; // consumer-only; never touched by offer()
+
+    /** Enqueues {@code node}. Safe from any number of concurrent threads. Zero allocation. */
+    void offer(WriteIntent node) {
+        node.setMpscNext(null);
+        WriteIntent prev = head.getAndSet(node);
+        prev.setMpscNext(node);
+    }
+
+    /**
+     * Dequeues the next intent, or {@code null} if the queue is empty <em>or</em> a producer is
+     * momentarily mid-{@link #offer} — see the class Javadoc. Single-consumer only.
+     */
+    WriteIntent poll() {
+        WriteIntent t = tail;
+        WriteIntent next = t.mpscNext();
+
+        if (t == stub) {
+            if (next == null) {
+                return null; // genuinely empty
+            }
+            tail = next;
+            t = next;
+            next = t.mpscNext();
+        }
+
+        if (next != null) {
+            tail = next;
+            return t;
+        }
+
+        WriteIntent h = head.get();
+        if (t != h) {
+            return null; // producer mid-offer; momentary, retry later
+        }
+
+        // t is the last real node and head hasn't moved past it: park the stub here so the
+        // next poll() (once a future offer() lands) has somewhere to advance from, then check
+        // whether t already gained a follower while we were doing this.
+        offer(stub);
+        next = t.mpscNext();
+        if (next != null) {
+            tail = next;
+            return t;
+        }
+        return null;
+    }
+
+    /** Cheap, conservative "might there be work" check — never a false negative, may be a false
+     *  positive (harmless: the caller just attempts a {@code tryLock()} that finds nothing). */
+    boolean hasWork() {
+        return head.get() != tail;
+    }
+}

flash/src/main/java/dev/relism/flash/http2/frame/Padding.java

@@ -0,0 +1,66 @@
+package dev.relism.flash.http2.frame;
+
+import dev.relism.flash.bytes.Pairs;
+import dev.relism.flash.http2.Http2ErrorCode;
+import dev.relism.flash.http2.Http2Exception;
+
+/**
+ * RFC 9113 §6.1 (DATA) / §6.2 (HEADERS) padding. When {@link FrameFlags#PADDED} is set, a
+ * frame's payload is laid out as: 1 pad-length byte, then the actual data (or header-block
+ * fragment), then that many padding bytes (RFC 9113 gives no meaning to the padding bytes
+ * themselves — they exist only to obscure payload size from network observers).
+ *
+ * <p>Padding is <b>not optional to support</b>: any client may send it on DATA or HEADERS
+ * regardless of whether the server ever sends padded frames itself.
+ *
+ * <h3>Flow control (forward note, not implemented here)</h3>
+ * RFC 9113 §6.9.1: padding bytes count against the DATA flow-control window even though they
+ * carry no data — the <em>whole</em> frame payload (pad-length byte + data + padding) is what a
+ * #dataLength(long)}. This class only locates the data range within the payload; it performs no
+ * flow-control accounting itself.
+ */
+public final class Padding {
+    private Padding() {}
+
+    /**
+     * Locates the actual data range within a payload that may or may not be padded. When
+     * {@code padded} is {@code false}, returns the whole payload unchanged (zero-cost — no
+     * padding byte to read, no arithmetic beyond the pack). When {@code true}, reads the
+     * pad-length byte at {@code buf[payloadOffset]}, validates it, and returns the data range
+     * that follows it.
+     *
+     * @return {@code Pairs.pack(dataOffset, dataLength)} — unpack with {@link Pairs#hi}/{@link Pairs#lo}
+     * @throws Http2Exception ({@code PROTOCOL_ERROR}) if {@code padded} is set but
+     *         {@code payloadLength == 0} (no room for the pad-length byte itself), or if the
+     *         claimed pad length is greater than or equal to the whole payload length (RFC 9113
+     *         §6.1: "If the length of the padding is the length of the frame payload or
+     *         greater, the recipient MUST treat this as a connection error")
+     */
+    public static long unpad(byte[] buf, int payloadOffset, int payloadLength, boolean padded) {
+        if (!padded) {
+            return Pairs.pack(payloadOffset, payloadLength);
+        }
+        if (payloadLength == 0) {
+            throw Http2Exception.of(Http2ErrorCode.PROTOCOL_ERROR,
+                    "PADDED flag set but the frame has no payload for the pad-length byte");
+        }
+        int padLength = buf[payloadOffset] & 0xFF;
+        if (padLength >= payloadLength) {
+            throw Http2Exception.of(Http2ErrorCode.PROTOCOL_ERROR,
+                    "pad length " + padLength + " >= frame payload length " + payloadLength);
+        }
+        int dataOffset = payloadOffset + 1;
+        int dataLength = payloadLength - 1 - padLength;
+        return Pairs.pack(dataOffset, dataLength);
+    }
+
+    /** Extracts the data offset from a value returned by {@link #unpad}. */
+    public static int dataOffset(long unpadded) {
+        return Pairs.hi(unpadded);
+    }
+
+    /** Extracts the data length from a value returned by {@link #unpad}. */
+    public static int dataLength(long unpadded) {
+        return Pairs.lo(unpadded);
+    }
+}

flash/src/main/java/dev/relism/flash/http2/frame/WriteIntent.java

@@ -0,0 +1,49 @@
+package dev.relism.flash.http2.frame;
+
+/**
+ * "Serialize yourself, then hand me the finished bytes." The interface a stream (and, eventually,
+ * connection-level singletons — the precompiled SETTINGS ACK, PING ACK, GOAWAY, WINDOW_UPDATE
+ * frames) implements to write through {@link Http2FrameWriter}.
+ *
+ * <h3>Layer 1 — serialize outside the lock</h3>
+ *
+ * By the time {@link Http2FrameWriter#write} is called, the implementation has already built its
+ * complete output (frame header + HPACK block + payload, or whatever the frame needs) into a buffer
+ * it owns — a per-stream scratch buffer, reused across writes, never allocated per call. {@link
+ * #buffer()}/{@link #offset()}/{@link #length()} just describe where that already-finished output
+ * lives. {@code Http2FrameWriter} never serializes anything itself; it only ever issues one bulk
+ * {@code write(buffer, offset, length)} while holding the connection's write lock — see {@code
+ * WRITER.md} for why that distinction is the entire point of this design (the lock must never be
+ * held across serialization work, only across the syscall).
+ *
+ * <h3>Intrusive queue linkage</h3>
+ *
+ * {@link #mpscNext()}/{@link #setMpscNext} are not part of the writer's public contract — they
+ * exist so a {@code WriteIntent} can double as an {@link IntrusiveMpscQueue} node with zero extra
+ * allocation when the writer is contended. Implementations provide simple field storage; nothing
+ * about the field is meaningful outside {@link IntrusiveMpscQueue}.
+ */
+public interface WriteIntent {
+
+  /** The buffer holding this intent's already-serialized bytes. */
+  byte[] buffer();
+
+  /** Offset of the first byte to write, within {@link #buffer()}. */
+  int offset();
+
+  /** Number of bytes to write, starting at {@link #offset()}. */
+  int length();
+
+  /** Intrusive MPSC queue linkage — see {@link IntrusiveMpscQueue}. Not for external use. */
+  WriteIntent mpscNext();
+
+  /** Intrusive MPSC queue linkage — see {@link IntrusiveMpscQueue}. Not for external use. */
+  void setMpscNext(WriteIntent next);
+
+  /**
+   * Called exactly once after this intent leaves the writer, whether the socket write succeeded or
+   * failed. Pooled control-frame intents use this hook to return their slot to the owning
+   * connection without allocating a completion object.
+   */
+  default void completed() {}
+}

flash/src/main/java/dev/relism/flash/http2/hpack/ContinuationAssembler.java

@@ -0,0 +1,84 @@
+package dev.relism.flash.http2.hpack;
+
+import dev.relism.flash.http2.Http2Exception;
+import dev.relism.flash.http2.Http2Limits;
+
+/** Reassembles one HEADERS/CONTINUATION sequence into a bounded contiguous connection buffer. */
+public final class ContinuationAssembler {
+  private final byte[] buffer;
+  private int streamId;
+  private int length;
+  private int continuationCount;
+  private boolean active;
+  private boolean complete;
+
+  public ContinuationAssembler() {
+    this(Http2Limits.MAX_HEADER_LIST_SIZE);
+  }
+
+  public ContinuationAssembler(int maximumBlockSize) {
+    if (maximumBlockSize <= 0) throw new IllegalArgumentException("non-positive block size");
+    buffer = new byte[maximumBlockSize];
+  }
+
+  public void begin(
+      int streamId, byte[] source, int offset, int fragmentLength, boolean endHeaders) {
+    if (active || streamId <= 0) throw Http2Exception.PROTOCOL_ERROR;
+    reset();
+    this.streamId = streamId;
+    append(source, offset, fragmentLength);
+    complete = endHeaders;
+    active = !endHeaders;
+  }
+
+  public void continuation(
+      int streamId, byte[] source, int offset, int fragmentLength, boolean endHeaders) {
+    if (!active || streamId != this.streamId) throw Http2Exception.PROTOCOL_ERROR;
+    if (++continuationCount > Http2Limits.MAX_CONTINUATION_FRAMES_PER_BLOCK) {
+      throw Http2Exception.PROTOCOL_ERROR;
+    }
+    append(source, offset, fragmentLength);
+    complete = endHeaders;
+    active = !endHeaders;
+  }
+
+  public byte[] buffer() {
+    return buffer;
+  }
+
+  public int length() {
+    return length;
+  }
+
+  public int streamId() {
+    return streamId;
+  }
+
+  public boolean isComplete() {
+    return complete;
+  }
+
+  public boolean isActive() {
+    return active;
+  }
+
+  public void reset() {
+    streamId = 0;
+    length = 0;
+    continuationCount = 0;
+    active = false;
+    complete = false;
+  }
+
+  private void append(byte[] source, int offset, int fragmentLength) {
+    if (source == null
+        || offset < 0
+        || fragmentLength < 0
+        || offset > source.length - fragmentLength
+        || fragmentLength > buffer.length - length) {
+      throw Http2Exception.COMPRESSION_ERROR;
+    }
+    System.arraycopy(source, offset, buffer, length, fragmentLength);
+    length += fragmentLength;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/hpack/HeaderListSizeException.java

@@ -0,0 +1,20 @@
+package dev.relism.flash.http2.hpack;
+
+/**
+ * Signals that a fully decoded HPACK block exceeded the configured header-list limit. The decoder
+ * delays this exception until the complete block has been consumed so dynamic-table state remains
+ * synchronized with the peer. The stream layer maps it to a request rejection without closing the
+ * HTTP/2 connection.
+ */
+public final class HeaderListSizeException extends RuntimeException {
+  private final long decodedSize;
+
+  HeaderListSizeException(long decodedSize) {
+    super("decoded header list exceeds limit: " + decodedSize, null, false, false);
+    this.decodedSize = decodedSize;
+  }
+
+  public long decodedSize() {
+    return decodedSize;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/hpack/HeaderSink.java

@@ -0,0 +1,13 @@
+package dev.relism.flash.http2.hpack;
+
+import dev.relism.fpr.core.ByteView;
+
+/** Receives decoded HPACK fields in wire order. */
+@FunctionalInterface
+public interface HeaderSink {
+  /**
+   * Accepts one field. The views are valid only for the duration of this call; a sink that needs
+   * them afterwards must copy them into storage owned by the stream.
+   */
+  void accept(ByteView name, ByteView value, boolean neverIndexed);
+}

flash/src/main/java/dev/relism/flash/http2/hpack/HpackDecoder.java

@@ -0,0 +1,156 @@
+package dev.relism.flash.http2.hpack;
+
+import dev.relism.flash.bytes.Pairs;
+import dev.relism.flash.bytes.PooledSlice;
+import dev.relism.flash.http2.Http2Exception;
+import dev.relism.flash.http2.Http2Limits;
+import dev.relism.fpr.core.ByteView;
+
+/** Stateful, allocation-free HPACK decoder for one HTTP/2 connection direction. */
+public final class HpackDecoder {
+  private final HpackDynamicTable dynamicTable;
+  private final int maximumHeaderListSize;
+  private final byte[] nameScratch = new byte[Http2Limits.MAX_HPACK_STRING_LENGTH];
+  private final byte[] valueScratch = new byte[Http2Limits.MAX_HPACK_STRING_LENGTH];
+  private final PooledSlice nameView = new PooledSlice();
+  private final PooledSlice valueView = new PooledSlice();
+
+  public HpackDecoder(int advertisedTableSize, int maximumHeaderListSize) {
+    if (maximumHeaderListSize < 0) throw new IllegalArgumentException("negative header-list size");
+    this.dynamicTable = new HpackDynamicTable(advertisedTableSize);
+    this.maximumHeaderListSize = maximumHeaderListSize;
+  }
+
+  public HpackDecoder() {
+    this(Http2Limits.HPACK_DYNAMIC_TABLE_SIZE_LOCAL, Http2Limits.MAX_HEADER_LIST_SIZE);
+  }
+
+  /** Decodes one complete header block. */
+  public void decode(byte[] buffer, int offset, int length, HeaderSink sink) {
+    if (buffer == null
+        || sink == null
+        || offset < 0
+        || length < 0
+        || offset > buffer.length - length) {
+      throw new IllegalArgumentException("invalid HPACK decode arguments");
+    }
+
+    int position = offset;
+    int limit = offset + length;
+    boolean sawHeader = false;
+    boolean oversized = false;
+    long headerListSize = 0;
+
+    while (position < limit) {
+      int first = buffer[position] & 0xff;
+      if ((first & 0x80) != 0) {
+        long decoded = HpackIntegers.decode(buffer, position, limit, 7);
+        int index = Pairs.hi(decoded);
+        position = Pairs.lo(decoded);
+        if (index == 0) throw Http2Exception.COMPRESSION_ERROR;
+        resolve(index, nameView, valueView);
+        sawHeader = true;
+        headerListSize += fieldSize(nameView, valueView);
+        if (headerListSize <= maximumHeaderListSize) sink.accept(nameView, valueView, false);
+        else oversized = true;
+        continue;
+      }
+
+      if ((first & 0x40) != 0) {
+        long decoded = HpackIntegers.decode(buffer, position, limit, 6);
+        int nameIndex = Pairs.hi(decoded);
+        position = Pairs.lo(decoded);
+        position = decodeName(buffer, position, limit, nameIndex);
+        position = decodeString(buffer, position, limit, valueScratch, valueView);
+        sawHeader = true;
+        headerListSize += fieldSize(nameView, valueView);
+        if (headerListSize <= maximumHeaderListSize) sink.accept(nameView, valueView, false);
+        else oversized = true;
+        dynamicTable.add(nameView, valueView);
+        continue;
+      }
+
+      if ((first & 0x20) != 0) {
+        if (sawHeader) throw Http2Exception.COMPRESSION_ERROR;
+        long decoded = HpackIntegers.decode(buffer, position, limit, 5);
+        dynamicTable.setMaximumSize(Pairs.hi(decoded));
+        position = Pairs.lo(decoded);
+        continue;
+      }
+
+      boolean neverIndexed = (first & 0x10) != 0;
+      long decoded = HpackIntegers.decode(buffer, position, limit, 4);
+      int nameIndex = Pairs.hi(decoded);
+      position = Pairs.lo(decoded);
+      position = decodeName(buffer, position, limit, nameIndex);
+      position = decodeString(buffer, position, limit, valueScratch, valueView);
+      sawHeader = true;
+      headerListSize += fieldSize(nameView, valueView);
+      if (headerListSize <= maximumHeaderListSize) sink.accept(nameView, valueView, neverIndexed);
+      else oversized = true;
+    }
+
+    if (oversized) throw new HeaderListSizeException(headerListSize);
+  }
+
+  public HpackDynamicTable dynamicTable() {
+    return dynamicTable;
+  }
+
+  private int decodeName(byte[] buffer, int position, int limit, int index) {
+    if (index != 0) {
+      resolveName(index, nameView);
+      return position;
+    }
+    return decodeString(buffer, position, limit, nameScratch, nameView);
+  }
+
+  private static int decodeString(
+      byte[] buffer, int position, int limit, byte[] scratch, PooledSlice output) {
+    if (position >= limit) throw Http2Exception.COMPRESSION_ERROR;
+    boolean huffman = (buffer[position] & 0x80) != 0;
+    long decoded = HpackIntegers.decode(buffer, position, limit, 7);
+    int encodedLength = Pairs.hi(decoded);
+    int dataStart = Pairs.lo(decoded);
+    if (encodedLength > limit - dataStart) throw Http2Exception.COMPRESSION_ERROR;
+    if (huffman) {
+      int decodedLength =
+          Huffman.decode(buffer, dataStart, encodedLength, scratch, 0, scratch.length);
+      output.reset(scratch, 0, decodedLength);
+    } else {
+      if (encodedLength > Http2Limits.MAX_HPACK_STRING_LENGTH)
+        throw Http2Exception.COMPRESSION_ERROR;
+      output.reset(buffer, dataStart, encodedLength);
+    }
+    return dataStart + encodedLength;
+  }
+
+  private void resolve(int index, PooledSlice name, PooledSlice value) {
+    if (index <= HpackStaticTable.LENGTH) {
+      byte[] staticName = HpackStaticTable.name(index);
+      byte[] staticValue = HpackStaticTable.value(index);
+      name.reset(staticName, 0, staticName.length);
+      value.reset(staticValue, 0, staticValue.length);
+      return;
+    }
+    dynamicTable.get(index - HpackStaticTable.LENGTH, name, value);
+  }
+
+  private void resolveName(int index, PooledSlice name) {
+    if (index <= 0) throw Http2Exception.COMPRESSION_ERROR;
+    if (index <= HpackStaticTable.LENGTH) {
+      byte[] staticName = HpackStaticTable.name(index);
+      name.reset(staticName, 0, staticName.length);
+      return;
+    }
+    dynamicTable.get(index - HpackStaticTable.LENGTH, name, valueView);
+    // An incremental-indexing representation can evict or compact the entry that supplied its
+    // indexed name. Preserve the name before insertion mutates the dynamic table arena.
+    System.arraycopy(name.array(), name.offset(), nameScratch, 0, name.length());
+    name.reset(nameScratch, 0, name.length());
+  }
+
+  private static long fieldSize(ByteView name, ByteView value) {
+    return (long) name.length() + value.length() + 32;
+  }
+}

flash/src/main/java/dev/relism/flash/http2/hpack/HpackDynamicTable.java

@@ -0,0 +1,132 @@
+package dev.relism.flash.http2.hpack;
+
+import dev.relism.flash.bytes.ArrayBackedByteView;
+import dev.relism.flash.bytes.PooledSlice;
+import dev.relism.flash.http2.Http2Exception;
+import dev.relism.fpr.core.ByteView;
+
+/**
+ * Per-connection HPACK dynamic table. Entries are kept in FIFO order in a descriptor ring while
+ * their bytes live in one bounded arena. The arena is compacted only when its free tail cannot hold
+ * the next entry, keeping every returned view contiguous.
+ */
+public final class HpackDynamicTable {
+  private final byte[] arena;
+  private final int[] nameOffsets;
+  private final int[] nameLengths;
+  private final int[] valueOffsets;
+  private final int[] valueLengths;
+  private final int advertisedMaximum;
+
+  private int maximumSize;
+  private int currentSize;
+  private int head;
+  private int count;
+  private int arenaEnd;
+
+  public HpackDynamicTable(int advertisedMaximum) {
+    if (advertisedMaximum < 0) throw new IllegalArgumentException("negative HPACK table size");
+    this.advertisedMaximum = advertisedMaximum;
+    this.maximumSize = advertisedMaximum;
+    this.arena = new byte[Math.max(1, advertisedMaximum)];
+    int entryCapacity = Math.max(1, advertisedMaximum / 32 + 1);
+    this.nameOffsets = new int[entryCapacity];
+    this.nameLengths = new int[entryCapacity];
+    this.valueOffsets = new int[entryCapacity];
+    this.valueLengths = new int[entryCapacity];
+  }
+
+  public int count() {
+    return count;
+  }
+
+  public int size() {
+    return currentSize;
+  }
+
+  public int maximumSize() {
+    return maximumSize;
+  }
+
+  /** Applies an RFC 7541 §4.2 table-size update and evicts oldest entries as necessary. */
+  public void setMaximumSize(int newMaximum) {
+    if (newMaximum < 0 || newMaximum > advertisedMaximum) throw Http2Exception.COMPRESSION_ERROR;
+    maximumSize = newMaximum;
+    evictToFit(0);
+    if (count == 0) arenaEnd = 0;
+  }
+
+  /** Inserts a new entry, copying its bytes before performing FIFO eviction. */
+  public void add(ByteView name, ByteView value) {
+    int byteLength = name.length() + value.length();
+    int entrySize = byteLength + 32;
+    if (entrySize > maximumSize) {
+      clear();
+      return;
+    }
+
+    evictToFit(entrySize);
+    if (arena.length - arenaEnd < byteLength) compact();
+
+    int slot = (head + count) % nameOffsets.length;
+    nameOffsets[slot] = arenaEnd;
+    nameLengths[slot] = name.length();
+    copy(name, arena, arenaEnd);
+    arenaEnd += name.length();
+    valueOffsets[slot] = arenaEnd;
+    valueLengths[slot] = value.length();
+    copy(value, arena, arenaEnd);
+    arenaEnd += value.length();
+    count++;
+    currentSize += entrySize;
+  }
+
+  /** Resolves a dynamic index where {@code 1} is the newest entry. */
+  public void get(int relativeIndex, PooledSlice name, PooledSlice value) {
+    if (relativeIndex < 1 || relativeIndex > count) throw Http2Exception.COMPRESSION_ERROR;
+    int slot = (head + count - relativeIndex) % nameOffsets.length;
+    name.reset(arena, nameOffsets[slot], nameLengths[slot]);
+    value.reset(arena, valueOffsets[slot], valueLengths[slot]);
+  }
+
+  public void clear() {
+    head = 0;
+    count = 0;
+    currentSize = 0;
+    arenaEnd = 0;
+  }
+
+  private void evictToFit(int incomingSize) {
+    while (count > 0 && currentSize + incomingSize > maximumSize) {
+      int slot = head;
+      currentSize -= nameLengths[slot] + valueLengths[slot] + 32;
+      head = (head + 1) % nameOffsets.length;
+      count--;
+    }
+    if (count == 0) arenaEnd = 0;
+  }
+
+  private void compact() {
+    int destination = 0;
+    for (int i = 0; i < count; i++) {
+      int slot = (head + i) % nameOffsets.length;
+      int nameLength = nameLengths[slot];
+      int valueLength = valueLengths[slot];
+      System.arraycopy(arena, nameOffsets[slot], arena, destination, nameLength);
+      nameOffsets[slot] = destination;
+      destination += nameLength;
+      System.arraycopy(arena, valueOffsets[slot], arena, destination, valueLength);
+      valueOffsets[slot] = destination;
+      destination += valueLength;
+    }
+    arenaEnd = destination;
+  }
+
+  private static void copy(ByteView source, byte[] target, int offset) {
+    if (source instanceof ArrayBackedByteView contiguous) {
+      System.arraycopy(contiguous.array(), contiguous.offset(), target, offset, source.length());
+      return;
+    }
+    for (int i = 0; i < source.length(); i++) target[offset + i] = source.byteAt(i);
+  }
+}