Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions quest/m1/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,8 @@ session outcomes, and embedder refusals inside session outcomes.
- [Untimed lite-07](/quest/m1/lite-untimed.md) - lite-07 carries an untimed track in both languages; lite-05/06 write send time
- [A watch and publish release ships assets()](/quest/m1/assets-release.md) - the release that lets the sites host the worklets
- [Dogfood hosted worklets](/quest/m1/dogfood-assets.md) - the moq.pro dashboard hosts the worklets and calls `assets()` after the release
- [TS pair acquisition](/quest/m1/ts-pair-acquisition.md) - fix the clean baseline and prove the grader's controls before #5328 enables its scheduled hard gate
- [Initial catalog routing](/quest/m1/initial-catalog-routing.md) - a usable fallback does not strand initial catalog acquisition with a spurious Unroutable
- [More tests under load](/quest/m1/test-flakes-2/README.md) - the second round of load-only failures, one quest per flake, fixed at the cause
- [END_OF_TRACK placement](/quest/m1/ietf-end-of-track-placement.md) - END_OF_TRACK rides the upstream's Location, and Rust's header stops claiming END_OF_GROUP
- [Plan: a public model](/quest/m1/model-public-surface.md) - every crate-private model item is made public or unnecessary, so a custom publisher or subscriber builds from the public model alone
Expand Down
62 changes: 62 additions & 0 deletions quest/m1/initial-catalog-routing.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,62 @@
# [M] Initial catalog acquisition survives a usable transport fallback

## Goal

Fix the initial `catalog.json` routing failure observed by TS validation,
with a deterministic regression and an automated round trip. A publisher
remaining available on a usable path must not leave the subscriber with a
spurious `Unroutable`; genuine refusal or disappearance still fails explicitly.

## Plan

Decided with the maintainer on 2026-10-11: reproduce and fix the cause, not
stop at diagnosis. This is independent of the pair grader's acquisition
contract. It is not a general transport redesign or a TS-specific retry path.

Local live, EIT and HRD runs during #5328 failed their initial `catalog.json`
track-info request shortly after selecting WebSocket fallback. The exporter
exited one with `moq: unroutable`; the publisher continued to normal EOF and
exited zero. The live grader received zero packets. Live and EIT ran
concurrently; HRD reproduced later. Do not count these as three independent
sequential reproductions. The full PR TS job passed, including EIT and HRD;
its event skipped the live and pair arms. No transport root cause is established.

Trace the failing initial request across publication discovery, catalog
availability, route selection and any session replacement. Preserve enough
session/broadcast/request identity to distinguish a real unavailable route
from a stale handle or handover race. First check current main and #5284:
that PR preserves a pending QUIC dial after an early WebSocket close, but
similar timing is not proof that it fixes this failure. If it owns the same
cause, add the missing regression there and verify the landed fix instead of
creating competing implementations. Otherwise keep the fix focused on the
layer that loses the usable route.

Turn the observed ordering into a controlled regression before changing
behavior. Hold or order the relevant events and mock time where needed;
do not rely on load, sleeps, repeated reruns or larger timeouts to reproduce
or hide the failure. Cover the successful fallback/acquisition path and the
adjacent terminal path so a fix cannot silently retry genuine refusal or
keep a disappeared publication alive. Preserve existing broadcast/track
identity and epoch semantics.

Verify the affected path with the TS round-trip harness and its existing
process-status, nonempty-content and decoder assertions. A successful rerun
alone does not finish the quest. Wire every regression and any new controls
into CI, with long live coverage at least nightly; retain diagnostic logs
and failing fixtures. Keep the publisher and exporter statuses in the verdict.

Keep the older short-duration capture regression and midstream `--stitch`
catalog-linger work separate unless reproduction establishes the same cause.
Update comments and affected feature docs with the fix. No new guide, public
API or wire change is expected; a newly required contract needs a maintainer
decision rather than an invented retry or compatibility layer.

## Related

- [Validation PR #5328](https://github.com/moq-dev/moq/pull/5328) - records the process failures and successful PR CI baseline
- [Closed fallback](/quest/m1/transport-upgrade/closed-fallback.md) - #5284 owns the pending QUIC dial after an early WebSocket close
- [Fallback PR #5284](https://github.com/moq-dev/moq/pull/5284) - check its actual effect on this reproduction before duplicating work
- [Transport upgrade](/quest/m1/transport-upgrade/README.md) - existing handover and epoch contracts
- [TS duration fidelity](/quest/m0/ts-duration-fidelity.md) - previously captured only a short suffix, not zero-packet initial acquisition
- [Replacement catalog linger](/quest/m1/ts-follow-catalog-bound.md) - bounds a midstream stitch, not initial discovery
- [Pair acquisition](/quest/m1/ts-pair-acquisition.md) - independent grader/mux finding
60 changes: 60 additions & 0 deletions quest/m1/ts-pair-acquisition.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
# [M] TS pair validation distinguishes acquisition from persistent misalignment

## Goal

The late-joining pair test accepts supported clean acquisition and detects
persistent table misalignment without mistaking interleave for emission time.
Its positive baseline and negative controls run in CI before the scheduled
hard gate in #5328 is merged.

## Plan

Decided with the maintainer on 2026-10-11: fix the finding, not just diagnose
it. Keep this separate from initial catalog routing and the broader redundant
TS coordinate/API planning in #5327. Fix the baseline before #5328 merges;
do not temporarily make the pair test non-gating.

The validation run in #5328 used `just test ts --pair`: two exports of one
45-second generated broadcast, the second joining five seconds late. Both
exporters and the publisher exited successfully. PAT/PMT agreement was 100%;
SDT had 20 common placements out of 23 in the union, or 86.96%, below the
existing 90% gate over 41 seconds of shared media. The mismatches were near
acquisition; 20 consecutive later SDT placements matched. These observations
do not establish permanent phase drift.

`test/ts/table-anchor.py` tags a table with the next PES PTS and infers a
timer-driven phase difference from median cadence. Check those assumptions
against the current slot scheduler and the capture's actual table placement.
Correct the grader, the mux, or both as the evidence requires; do not assume
the exporter is defective because that diagnostic says so.

Define and document the acquisition behavior being graded for the generated
profile. Any distinction between acquisition tables and recurring tables must
follow an observable stream condition with automated evidence. Do not ignore
an arbitrary number of seconds, reduce the agreement or coverage thresholds,
or shrink the shared window to the surviving table emissions. Keep missing
expected tables and inadequate evidence distinguishable from a passing check.

Add deterministic controls for supported late acquisition and changed PES
interleave, plus persistent phase offset, drift and table loss within shared
media. The controls must fail for their intended reasons and must prevent a
grader that accepts everything from passing. Keep table carriage, cadence and
packet identity separate: this quest does not promise byte-identical TS legs,
introduce `--sync`, or claim ST 2022-7 compatibility.

Run the short controls on relevant PRs and the real pair arm at least nightly,
using shared `just test` recipes. Preserve captures and reports on failure.
Require a passing clean pair baseline and failing negative controls before
unblocking #5328; a green PR run that skips the pair arm is not that evidence.
Update `test/ts/README.md` and misleading grader comments/diagnoses inline.

Public API and wire: none expected. New shared-coordinate or synchronization
contracts remain with the broader alignment planning.

## Related

- [Validation PR #5328](https://github.com/moq-dev/moq/pull/5328) - its scheduled pair gate must wait for this fix
- [Pair gate review](https://github.com/moq-dev/moq/pull/5328#discussion_r4243103700) - a known-failing baseline would generate daily alerts
- [TS roadmap #5327](https://github.com/moq-dev/moq/pull/5327) - owns the broader redundant alignment planning
- [Deterministic TS legs](/quest/m2/ts-hitless.md) - packet identity requires a separate contract and proof
- [Initial catalog routing](/quest/m1/initial-catalog-routing.md) - distinct acquisition failure found by the same validation work
Loading