fix(net): write IETF data stream headers in one write

js/net/src/ietf/fetch.ts

@@ -1,6 +1,6 @@
 import type * as Epoch from "../epoch.ts";
 import type * as Path from "../path.ts";
-import type { Reader, Writer } from "../stream.ts";
+import { type Reader, Writer } from "../stream.ts";
 import { hasRangeFilters, isDraft20 } from "./filter.ts";
 import * as Message from "./message.ts";
 import * as Namespace from "./namespace.ts";
@@ -22,9 +22,14 @@ export class FetchHeader {
 		this.requestId = requestId;
 	}
 
-	/** Write the header, which the stream type must already precede. */
-	async encode(w: Writer, _version: IetfVersion): Promise<void> {
-		await w.u62(this.requestId);
+	/** Write the stream type and the header in one write, like a subgroup header. */
+	async encode(w: Writer, version: IetfVersion): Promise<void> {
+		await w.write(
+			await Writer.collect(version, async (h) => {
+				await h.u53(FetchHeader.type);
+				await h.u62(this.requestId);
+			}),
+		);
 	}
 
 	/** Read the header, with the stream type already consumed. */

js/net/src/ietf/object.test.ts

@@ -1,7 +1,8 @@
 import { expect, test } from "bun:test";
-import { Cursor } from "../stream.ts";
+import { Cursor, Reader, Writer } from "../stream.ts";
 import { Timescale, Timestamp } from "../time.ts";
-import { decodeObjectTime, encodeObjectExtensions } from "./object.ts";
+import { FetchHeader } from "./fetch.ts";
+import { decodeObjectTime, encodeObjectExtensions, Group } from "./object.ts";
 import { Version } from "./version.ts";
 
 const VERSION = Version.DRAFT_17;
@@ -30,3 +31,40 @@ test("a conversion past what a Timestamp holds is refused before anything is wri
 		RangeError,
 	);
 });
+
+test("a data stream's header goes out in one write, stream type included", async () => {
+	for (const version of [Version.DRAFT_14, Version.DRAFT_22]) {
+		const chunks: Uint8Array[] = [];
+		const writer = new Writer(
+			new WritableStream<Uint8Array>({ write: (chunk) => void chunks.push(new Uint8Array(chunk)) }),
+			version,
+		);
+
+		// Every optional field present, with values past a one-byte varint.
+		const group = new Group({
+			trackAlias: 1_000_000n,
+			groupId: 70_000,
+			subGroupId: 300,
+			publisherPriority: 200,
+			flags: {
+				hasExtensions: true,
+				hasSubgroup: true,
+				hasSubgroupObject: false,
+				hasEnd: true,
+				hasPriority: true,
+				firstObject: true,
+			},
+		});
+		await group.encode(writer, version);
+		await new FetchHeader({ requestId: 1_000_000n }).encode(writer, version);
+		expect(chunks.length).toBe(2);
+
+		const decoded = await Group.decode(new Reader(undefined, chunks[0], version), version);
+		expect(decoded).toEqual(group);
+
+		const fetch = new Reader(undefined, chunks[1], version);
+		expect(await fetch.u53()).toBe(FetchHeader.type);
+		expect((await FetchHeader.decode(fetch, version)).requestId).toBe(1_000_000n);
+		expect(await fetch.done()).toBe(true);
+	}
+});

js/net/src/ietf/object.ts

@@ -161,7 +161,15 @@ export class Group {
 		this.publisherPriority = publisherPriority;
 	}
 
+	/**
+	 * Write the header, stream type first, in one write: a peer that classifies a stream from
+	 * its first read (imquic) drops a header split across several.
+	 */
 	async encode(w: Writer, version: IetfVersion): Promise<void> {
+		await w.write(await Writer.collect(version, (h) => this.#encode(h, version)));
+	}
+
+	async #encode(w: Writer, version: IetfVersion): Promise<void> {
 		if (!this.flags.hasSubgroup && this.subGroupId !== 0) {
 			throw new Error(`Subgroup ID must be 0 if hasSubgroup is false: ${this.subGroupId}`);
 		}

js/net/src/ietf/publisher.ts

@@ -1336,7 +1336,6 @@ export class Publisher {
 			streams.opened += 1;
 			if (sequence !== undefined) ranking.add(stream, sequence, object);
 
-			await stream.u53(FetchHeader.type);
 			await new FetchHeader({ requestId }).encode(stream, version);
 
 			if (fill.kind !== "group") {

quest/m1/audit-interop.md

@@ -86,7 +86,11 @@ Waiting on others:
   (closes on a request before SETUP completes),
   [meetecho/imquic#42](https://github.com/meetecho/imquic/issues/42) and
   [Eyevinn/moqlivemock#157](https://github.com/Eyevinn/moqlivemock/issues/157)
-  (interop clients drop the URL path), plus moxygen's filter type 250
+  (interop clients drop the URL path),
+  [meetecho/imquic#43](https://github.com/meetecho/imquic/issues/43)
+  (drops a data stream whose header spans reads; owned by
+  [JS subgroup data through imquic](/quest/m1/interop-coverage/js-imquic-objects.md)),
+  plus moxygen's filter type 250
   ([facebookexperimental/moxygen#225](https://github.com/facebookexperimental/moxygen/issues/225)),
   which also fails moqx's client against our relay.
 - The JS publisher refuses every FETCH, so every relay that fills a late

quest/m1/interop-coverage/README.md

@@ -41,7 +41,7 @@ coverage on `main` at `3ab6cbf81d`.
 ## Required
 
 - [JS namespace lifecycle](/quest/m1/interop-coverage/js-namespace-done.md) - moq-dev-js subscribe-namespace-lifecycle sees NAMESPACE_DONE with publisher and subscriber in one process
-- [JS subgroup data through imquic](/quest/m1/interop-coverage/js-imquic-objects.md) - moq-dev-js data-subgroup-basic receives objects from imquic over WebTransport
+- [JS subgroup data through imquic](/quest/m1/interop-coverage/js-imquic-objects.md) - moq-dev-js data-subgroup-basic passes against imquic, or the cell is classified with imquic's split-header bug reported upstream
 - [Late join and FETCH](/quest/m1/interop-coverage/late-join.md) - a late subscriber gets the current group from object 0, and a standalone FETCH returns a past group exactly
 
 ## Related

quest/m1/interop-coverage/js-imquic-objects.md

@@ -1,18 +1,42 @@
-# [S] JS subgroup data reaches the subscriber through imquic
+# [XS] JS subgroup data reaches the subscriber through imquic
 
 ## Goal
 
 The runner's moq-dev-js `data-subgroup-basic` case passes against imquic
-(draft-22, WebTransport), where today the subscriber receives no objects.
-If the cause is in imquic, it is reported upstream and the cell is
-classified instead.
+(draft-22, WebTransport), or the cell is classified with the imquic issue
+reported upstream. Check the cell on the runner's report.
 
 ## Plan
 
-Found in the runs for
-[englishm/moq-interop-runner#173](https://github.com/englishm/moq-interop-runner/pull/173)
-on 2026-10-22. moq-dev-rs passes against imquic on the same draft, so
-compare what the two clients put on the wire (transport, SUBSCRIBE
-fields, stream types) before suspecting imquic.
+imquic drops a SUBGROUP_HEADER (or FETCH_HEADER) that arrives across more
+than one stream read: `imquic_moq_parse_message` creates the data stream on
+the first chunk but only buffers bytes once the header has parsed, so the
+rest of the header and every object are discarded, and the relay logs
+`No track found for that subscription` with alias 0 at the stream's FIN.
+QUIC may split a stream anywhere, so this is imquic's bug, reported as
+[meetecho/imquic#43](https://github.com/meetecho/imquic/issues/43).
+`@moq/net` 0.4.4 wrote the header one field per write, which the polyfill
+sent as separate frames; moq-dev-rs buffers it and passes.
 
-Public API: unknown until the cause is found. Wire: none expected.
+[moq-dev/moq#6504](https://github.com/moq-dev/moq/pull/6504) writes both
+data stream headers in one write. The maintainer kept it on 2026-10-23
+even though the root cause is imquic's: it matches the Rust publisher and
+the one-write control messages from
+[moq-dev/moq#6255](https://github.com/moq-dev/moq/pull/6255), and also
+covers FETCH_HEADER. Reproduced and verified against the runner's imquic
+relay image and `lminiero.it:9000`: 0.4.4 fails every run, and passes once
+patched.
+
+Remaining:
+
+- The cell turns green once imquic buffers a split header, or once the
+  runner's JS client moves to a `@moq/net` release with
+  [moq-dev/moq#6504](https://github.com/moq-dev/moq/pull/6504). Delete
+  this quest then.
+- On `main`, `request(path, { announced: true })` asks SUBSCRIBE_NAMESPACE
+  for the tuple's own namespace, and imquic never answers with the shorter
+  routing prefix the publisher announced, so the subscriber never
+  subscribes. 0.4.4 asked for the prefix. The runner's client hits this
+  when it moves past 0.4.4 unless it requests blindly.
+
+Public API: none. Wire: none (bytes unchanged, fewer writes).