Draft-20 backlog: documentation cleanup and relay refusal codes

[floatdrop]

This PR clears two items from the draft-20 backlog: documentation cleanup and relay refusal-code fixes. There is one commit per item. Each was reviewed by moqt-reviewer, and each code change has tests that were seen failing first, either before the fix or under a mutant that removes the behaviour.

Documentation

• e22125d STATUS.md and the namespace handle docs now match the code and draft-20:
  • The Limitations entry saying duplicates aren't compared is gone, because the relay does compare them.
  • LOC PropAudioLevel is 0x0C, not 0x0A.
  • The "handles the application reads itself" note now separates role checks from scope checks, and names the handles that have no broker.
  • The request-stream GOAWAY note now says such a GOAWAY is checked and otherwise ignored.
  • The FILL_PARAMETERS row is now DONE, and the INCLUDE_PROPERTIES row points at the Group Order draft gap.
  • The Joining FETCH mention (removed in draft-20) is dropped.
  • NamespacePublication / IncomingNamespacePublication docs: NAMESPACE / NAMESPACE_DONE answer only a SUBSCRIBE_NAMESPACE (§10.17, §10.18). A namespace is withdrawn by cancelling the request, not by a FIN (§6.2, §3.3.2). The docs now also name the broker settings a reader needs.
• 9a2c992 corrects 13 stale § citations in code comments, each checked against the draft text: padding, GREASE, datagram flags, fetch ordering, caching, draining, and relaynet ALPN. Three of the old review's suggested corrections were themselves wrong and were left as they were.
• d12096b records one interpretation: a datagram is filtered as Subgroup 0 under SUBGROUP_FILTER (§2.2, §5.1.4).

Relay

• 33dd402 When an upstream answers a SUBSCRIBE with GOING_AWAY, the subscriber now gets GOING_AWAY with its Retry Interval, instead of INTERNAL_ERROR. It also counts as "no publisher yet" for RENDEZVOUS_TIMEOUT.
• 38f0a53, dbaaa73 A SUBSCRIBE's refusal no longer depends on the order candidates answer in:
  • Ranking: unknown Mandatory Track Property, then malformed Track Properties, then any other refusal, then GOING_AWAY, then TIMEOUT, then DOES_NOT_EXIST or a transport failure.
  • Ties go to the soonest retry, then to a specific code over INTERNAL_ERROR.
  • Retry 0: an upstream GOING_AWAY with Retry Interval 0 gets the relay's own ~1s instead.
  • Transport failures don't end a RENDEZVOUS_TIMEOUT hold, and that candidate is asked again on the next look. Lack of stream credit is excluded: the hold ends on it as before.
• 5e61d8a TRACK_STATUS is forwarded upstream when the track has no Established subscription (§10.15 MAY):
  • Every candidate SUBSCRIBE would try is asked, concurrently.
  • Answers combine as the relay's SUBSCRIBE_OK would: the entry's Track Properties, else the first answer's, and the largest LARGEST_OBJECT.
  • If none answers OK, the subscriber gets the refusal SUBSCRIBE would give.
  • Forwarding counts against MaxSubscriptionsPerSession (§13.1) and is bounded at 5s (§13.6).
  • Loop guard as for SUBSCRIBE and stitch FETCH.

Decisions and interpretations

• Maintainer decisions:
  • forward TRACK_STATUS, asking every candidate and taking the maximum;
  • cap plus 5s timeout;
  • precedence GOING_AWAY > TIMEOUT > DOES_NOT_EXIST;
  • pass an upstream GOING_AWAY through, with 0 replaced by the relay's retry;
  • soonest retry wins ties;
  • transport failures count as "no publisher yet".
• Interpretation, marked in the code: GOING_AWAY outranks §10.2.6's DOES_NOT_EXIST for a known publisher that is draining.
• Behaviour changes:
  • A namespace-only TRACK_STATUS no longer gets an immediate empty OK. TestTrackStatus_ReplyEmptyPropertiesForKnownNamespace pinned that and is removed.
  • TestSubscribe_UpstreamRejects_PropagatesRejection's GOING_AWAY row changes from INTERNAL_ERROR to GOING_AWAY.

Known gaps (added to STATUS.md)

• Concurrent TRACK_STATUS requests for one track aren't coalesced.
• A SUBSCRIBE the relay can't open for lack of stream credit is answered DOES_NOT_EXIST with no retry; EXCESSIVE_LOAD would fit better (§10.6.2).

Verification

• go test ./... and golangci-lint run pass.
• go test -race ./pkg/relay/... passes.

🤖 Generated with Claude Code