[kixelated]

Problem

On release, a cache-miss FETCH can outrun the subscription's joining fill. The fetched group then occupies the cache, the live tail is refused as a duplicate, and live readers wait for an end that never arrives.

Approach

Hand-port the settled relay fix from moq-dev/moq#7135. Only the group named by SUBSCRIBE_OK waits for its own head. The cache answers an arriving head; a settled fill with a remaining miss permits a FETCH. Reader cancellation and track termination release the wait. The PUBLISH_DONE grace also polls queued misses, so a head that settles there can release its FETCH.

The subscriber coverage uses release's paused Tokio harness. No main-only range-fill or subscription-widening machinery is ported because release has neither. Main quest bookkeeping stays in moq-dev/moq#7669.

Impact

• Public API: unchanged; the request completion helper is crate-private.
• Wire: unchanged; this suppresses redundant FETCH requests without changing their encoding.
• No package version changes, tags, or publication.

Alternatives

Cherry-picking would import main's incompatible track-fill model. Delaying all cache misses would stall older history behind an unrelated head, so only the joined group waits.

Validation

• Reproduced on unfixed release: both joined-head miss tests emitted an extra FETCH; the older-history control passed.
• All 13 joining-fetch tests pass, including six added cases for waiting, fallback, cache delivery, older history, cancellation, track abort, and PUBLISH_DONE grace.
• The strict release CI test half passes: 5,188 default tests, 387 capture tests, and 133 CLI play/capture tests.
• cargo semver-checks check-release -p moq-net --baseline-rev origin/release --release-type patch --all-features: 229 checks pass; no semver update required.
• The strict release CI check half passes, including WASM, selected device configurations, and repository lints. Checks run through release's Nix shell with BASH_ENV unset inside it, so a cached Codex shell snapshot cannot replace the pinned toolchain.
• After merging origin/release (d6ffbbb69b) conflict-free into c59489c5da: cargo clippy -p moq-net --all-targets --all-features -D warnings is clean, and cargo test -p moq-net --all-features passes 1,613 tests, including all 13 joining-fetch tests.

Review

• Head c59489c5da only adds a conflict-free merge of origin/release. git diff origin/release...c59489c5da is byte-identical to git diff origin/release...87e8ef9c8e, so the round 2 record below covers it.
• Latest: Claude Code (quest review), Claude Opus 5.5, medium effort, round 2 against origin/release, reviewed 87e8ef9c8e: no findings.
• Round 1 dispositions, accepted in round 2:
  • Missing-head fallback: refuted. The settled design deliberately leaves an absent head costing only its own group, until its reader cancels or the subscription ends; see the original review response.
  • Older-history ownership redesign: refuted for this backport. Older history must FETCH immediately, as the quest Plan specifies. The review's Absolute-join collision is pre-existing and degrades a later join rather than causing the joined-head live-reader hang. It is reported as a suggested main follow-up.
  • PUBLISH_DONE grace: fixed by polling queued misses during settlement. The new paused-time test reproduced the missing FETCH before this fix, and now passes.

Main-first prerequisite

Cleared 2026-11-05 by moq-dev/moq#7903, merged as edc7ef73ca. It adds the same group_fetches.poll(waiter) to main's PUBLISH_DONE grace loop, just before poll_settled, matching this PR's poll. It also adds a_miss_fetches_during_the_publish_done_grace, which failed on unfixed main. Its other change drives a lingering copy's fetches during resumed SUBSCRIBE setup. That has no release counterpart, because release creates its fetch set after SUBSCRIBE_OK and never lingers.

Follow-ups

• The main quest (quest/m0/backport-fill-head-race.md) keeps the maintainer's merge into release and the moq-net patch publication.
• Suggested follow-up quest: verify and fix the pre-existing collision between an Absolute joining fill and an older-history FETCH on main. The local reviewer identified a possible degraded join; the joined-largest-group hang is covered here.
• Main bookkeeping also records a test-discovery follow-up: Bun filename filters discovered the API checker's generated release source tree while checking main.
• Suggested follow-up quest: prevent a cached Codex shell snapshot from overriding an explicit Nix shell after switching branches. The cached toolchain lacked the WASM target; the declared release toolchain has it.

(Written by GPT-6.1 Sol and Claude Opus 5.5)

[kixelated]

Automated review by review (OpenAI)

Reviewed commit c59489c5da6ead7395894ec11d1d821639d1c44f against release base d6ffbbb69b29e492734bd04da07efe3c3666c8ea. No earlier published review was present on this PR.

No actionable findings. Direction: keep this narrow backport. Waiting only for the named joined group addresses the competing-producer race without importing main's range-fill machinery or blocking unrelated history. The cancellation and unrelated-history concerns from #7135 are addressed here: the wait observes reader withdrawal and terminal track state, and cache resolution retires the request without rejecting its readers. The added PUBLISH_DONE poll keeps already queued misses moving during settlement and matches merged #7903.

The new helper remains crate-private, with no wire-format change. Tests cover waiting/fallback, cache delivery, older-history progress, cancellation, track abort, and grace-period progress.

Verification: GitHub-only static review of the changes, surrounding lifecycle/cache code, tests and prior findings, with an independent cancellation/locking cross-check. Tests and interoperability were not run; the new head-arrival test injects internal fill state. Author-reported local results were not independently reproduced. GitHub Check/Test are still running; an earlier release run has a failed gate following a cancelled Scope job.

(Written by OpenAI)

[pullfrog]

✅ No new issues found.

Reviewed changes

Backport of the joined-head FETCH race fix (#7135, with #7903's PUBLISH_DONE poll) onto release at c59489c5da, including the conflict-free merge of origin/release (d6ffbbb69b). The diff matches the claimed scope exactly: two files, no stray changes from the merge.

• Joined-head gate — wait_group_head parks a cache-miss FETCH for the group named by SUBSCRIBE_OK's Largest until the subscription's own fill settles, then re-checks the cache and only runs run_group_fetch for a remaining miss. Older history, a missing subscription, a settled fill, and a departed reader all fall through without waiting.
• Request resolution — new crate-private group::Request::poll_resolved reports when the cache or a terminal track state already answers the request and retires the attempt without writing a rejection, so readers that resolve through the track are not failed.
• PUBLISH_DONE grace — the grace loop now polls group_fetches before poll_settled, so a head that settles during settlement can release its FETCH instead of waiting for teardown.
• Tests — JoinRun::subscribe factored out of start, plus six paused-time cases covering the wait, cache delivery, older history, reader cancellation, track abort, and the grace.

I re-ran the suite locally and mutation-checked the two load-bearing hunks: cargo test -p moq-net --all-features --lib joining_fetch passes 26 tests; returning Some(request) immediately from wait_group_head fails five of the new cases (the older-history control still passes), and deleting the new group_fetches.poll(waiter) from the grace loop fails the grace case.

The outstanding behavior — a fill that never settles parks that one group's miss until the reader cancels or the subscription ends — is the design settled on main, and the Absolute-join collision below Largest remains pre-existing with a follow-up already recorded. No new evidence to reopen either.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash-1017 (free via Pullfrog for OSS) ｜ 𝕏}