fix(moq): reconcile publisher history construction

[allyblockcast]

Summary

• add TrackWriter::subgroups_with_history(NonZeroU64) on main while preserving shared TrackState propagation into SUBSCRIBE_OK
• migrate MMTP source and repair publishers to construction-time history configuration
• preserve bounded history and u64::MAX semantics while accepting valid out-of-order relay subgroups
• report the largest retained (group_id, object_id) for subscription/resume state even when creation order differs

Head / Base

• base: main
• head: 3ae50d6d4b314b05bd8cc825d9f1ca3b17f97258

Compatibility

• keeps the publisher-to-relay mirror bound from BLO-10339
• keeps downstream relay re-advertisement from BLO-10419
• uses immutable creation cursors so non-prefix pruning cannot shift a relay reader cursor
• preserves sequence-ordered resume locations independently of creation-order delivery
• leaves datagram construction on the existing shared-state setter because this follow-up is scoped to subgroup/MMTP publishers

Verification

• cargo test -p moq-transport -p moq-pub-mmtp: 280 + 73 unit tests passed; integration/example tests also passed
• history_window_accepts_out_of_order_groups_and_prunes_by_id: sequence 10, 9, 8, 11, window 3, retains 10, 9, 11
• latest_uses_largest_group_after_out_of_order_create: group 11 followed by late group 10 still reports (11, 0)
• SUBSCRIBE_OK propagation regressions from BLO-10339 and BLO-10419 remain green in the 280-test transport run
• cargo fmt --all -- --check: passed
• cargo clippy -p moq-transport -p moq-pub-mmtp -- -D warnings: passed
• cargo clippy ... --all-targets -- -D warnings: blocked by pre-existing drop_non_drop in moq-transport/src/session/subscriber.rs:1899

Risks

• retained subgroup latest-location lookup is linear in the bounded history queue; the queue is already traversed for duplicate detection and pruning
• unbounded subgroup streams retain creation-order delivery while latest-location lookup remains sequence ordered

Paperclip: https://paperclip.blockcast.net/BLO/issues/BLO-23538

[allyblockcast]

🔗 Paperclip issue: BLO-23538
🔗 Paperclip issue: BLO-10339
🔗 Paperclip issue: BLO-10419

[allyblockcast]

🔗 Paperclip issue: BLO-23538
🔗 Paperclip issue: BLO-10339
🔗 Paperclip issue: BLO-10419

[allyblockcast]

@ally please review head b41650d, focusing on preservation of shared TrackState/SUBSCRIBE_OK propagation and the bounded-group overflow/order semantics.

[allyblockcast]

Ally — Consolidated PR Review

Lenses: pr-review-toolkit (code, tests, comments, errors, types) + gstack/review + native-codex.
Reviewed head: b41650d

Critical Issues (0)

Important Issues (1)

• [gstack/review] moq-transport/src/serve/subgroup.rs:139 — Bounded tracks now reject a valid out-of-order subgroup instead of retaining it. Subgroup explicitly allows group IDs to be received out of order (subgroup.rs:308), and a subscriber applies the advertised history window to its relay mirror before passing inbound subgroup headers to this writer. Thus a peer sending group 10 followed by group 9 causes ServeError::Duplicate and aborts receive processing, losing valid data. The new test codifies that regression.
  • Preserve out-of-order acceptance for received subgroups and make pruning independent of insertion order, or limit this monotonicity rule to the local MMTP publisher path.

Suggestions (0)

Strengths

• subgroups_with_history keeps history configuration and SUBSCRIBE_OK propagation coupled at construction time.
• The u64::MAX pruning case has targeted regression coverage.

Recommended Action

1. Address the Important issue before merge.

[allyblockcast]

@ally please re-review head 8ef99a2, focusing on out-of-order subgroup acceptance, non-prefix history pruning, and reader cursor stability.

[allyblockcast]

Ally — Consolidated PR Review

Lenses: pr-review-toolkit (code, tests, comments, errors, types) + gstack/review + native-codex.
Reviewed head: 8ef99a2

Critical Issues (0)

Important Issues (1)

• [native-codex] moq-transport/src/serve/subgroup.rs:252 — still returns the last-created retained subgroup, not the largest . Out-of-order acceptance now makes insertion order diverge from sequence order: after receiving group 11 and then a valid late group 10, reports group 10. That location is used for subscription/resume state and can make a peer start from stale history or omit group 11.
  • Select the maximum retained location by group/object ordering (rather than ), and add a regression that appends an older group after the newest one.

Suggestions (0)

Strengths

• Immutable creation indexes correctly avoid shifting reader cursors when history pruning removes a non-prefix entry.
• The pruning predicate handles safely and the new tests exercise the intended out-of-order retention window.

Recommended Action

1. Fix the Important issue this cycle before relying on out-of-order subgroup support for resume locations.

[allyblockcast]

@ally please re-review head 3ae50d6, focusing on sequence-ordered largest_location reporting after out-of-order subgroup creation and the new regression coverage.

[allyblockcast]

Ally — Consolidated PR Review

Lenses: pr-review-toolkit (code, tests, comments, errors, types) + gstack/review + native-codex.
Reviewed head: 3ae50d6

Prior Findings Dispositioned (2)

• prior:b41650d important 1 — fixed — moq-transport/src/serve/subgroup.rs:192 — pruning now retains entries through VecDeque::retain across the complete creation-order queue, rather than assuming out-of-order subgroup groups form a removable prefix.
• prior:8ef99a2 important 1 — fixed — moq-transport/src/serve/subgroup.rs:238 — latest() now takes the maximum retained (group_id, object_id) instead of using the final creation-order entry.

Critical Issues (0)

Important Issues (0)

Suggestions (0)

Strengths

• subgroups_with_history atomically persists the history window before transitioning the track, preserving SUBSCRIBE_OK propagation.
• Regression coverage exercises both out-of-order retention and the latest-location ordering case.

Recommended Action

1. No blocking issues found in this revision.