feat(net)!: an undeclared Rust timescale means untimed

[kixelated]

Problem

moq_net::track::Info::default() (and so create_track(name, None) and accept(None)) silently gave a track a millisecond timescale, while @moq/net treats an omitted timescale as untimed (#4968). The publish-timestamp and lite-07 untimed quests need each track to say whether it is timed before publishers stop inventing timestamps.

Approach

• moq-net: removed impl Default for Timescale; Info::default() now has timescale: None. The lite encoder (TRACK_INFO, frame and datagram send times) and the lite datagram/fetch decoders name Timescale::MILLI for an untimed track, as does moq-audio's decode quantum floor.
• Every Rust publisher that stamps frames declares milliseconds explicitly: hang catalog (Catalog::default_track_info, also used for MSF now), moq-mux timeline, TS SI tables (ts/si.rs, found by the moq-cli import/export tests), moq-boy status, moq-room chat, moq-stats JSON tracks, moq-bench, moq-c data_track, moq-ffi json/flate tracks, the moq-tokio and moq-json examples, and the moq-mux doc examples. Media tracks already pin their scale (hang::container::track_info, moq-c/moq-ffi raw tracks at microseconds).
• Tests that write timestamps on a track created with None now use a millisecond track (Info::timed() inside moq-net, a local timed() helper elsewhere), preserving what they tested before. New regression test an_undeclared_timescale_is_untimed; broadcast_moq_lite_05_default_timescale became broadcast_moq_lite_05_untimed_track.
• Docs: doc/setup/upgrade.md (Unreleased), doc/lib/js/net.md, rustdoc on Info, Info::with_timescale, create_track, hang::container::track_info. Quest file deleted and references updated; publish-timestamp and lite-untimed are now unblocked.
• Follow-up on this branch: writers that still stamped a frame after create_track(name, None) now declare Timescale::MILLI. That includes the io_uring session and WebTransport tests, the relay io_uring runtime tests, and the benches and loom tests that write timestamps. Tracks that never stamp a frame stay untimed. Merged origin/main (through the C++ package landing).

Decision

CI failed because lite_session_over_the_worker writes Timestamp::ZERO on a track whose timescale is now untimed.

• ✅ Keep the TimestampMismatch refusal. These are transport and bench paths, not untimed-track tests, so the track gets an explicit millisecond timescale, which None used to mean.
• Skip the test, or accept a timestamp on an untimed track. Rejected.

Left for the maintainer, unchanged:

1. Convenience constructor for timed tracks. Recommended: keep the with_timescale call; no new public API.
2. The stale rs/moq-mux/AGENTS.md line. Agent files stay untouched unless a maintainer asks.
3. TS SI tables stay on milliseconds. Settle a finer scale with the shared-clock quest.

Impact

• Breaking (Rust API): moq_net::Timescale no longer implements Default.
• Breaking (Rust behavior): track::Info::default(), broadcast::Producer::create_track(name, None), and track::Request::accept(None) create an untimed track; writing a timestamp to it fails with TimestampMismatch. Declare one with Info::default().with_timescale(Timescale::MILLI).
• This follow-up adds no public API. The millisecond helpers are private to the tests and benches.
• moq-ffi / moq-c / bindings: no API change; their data tracks stay millisecond-timed and raw/media tracks stay microsecond.
• Wire: none. An untimed track on lite-05+ still declares milliseconds and carries send times, as before.

Alternatives

• A public convenience such as track::Info::timed(scale) or a Timescale-to-Info conversion, so timed publishers don't write Info::default().with_timescale(...). Not added: no consumer needed new API to land this, and it can come with the publish-timestamp quest if wanted.
• Rewriting tests that use Timestamp::ZERO as a placeholder to write untimed frames instead. Rejected: many depend on timestamps (expiry, retention), and a timed track keeps each test's original meaning.

Follow-ups

• rs/moq-mux/AGENTS.md still says Info::default() is milliseconds; left alone since agent files need maintainer prompting.
• TS SI tables keep the old millisecond scale; whether they should share the importer's microsecond scale belongs with the shared-clock work.

Tested

• cargo nextest run --locked -p moq-uring --no-fail-fast --test session --test web -j 1: 13 passed, including lite_session_over_the_worker.
• cargo nextest run --locked -p moq-relay --features io-uring --no-fail-fast --test runtime_uring -j 1: 6 passed.
• quest check: 418 documents ok, after merging origin/main.
• just check: passed on the timestamp fixes (5615 + 458 + 157 + 97 Rust tests, JS, Python). The later origin/main merge did not touch those writers.

(Written by Grok 4.7)

[kixelated]

Quest outcome: implemented quest/m1/rust-untimed-default.md in full; left as a draft for the maintainer's decisions below.

Open decisions:

1. Convenience constructor for timed tracks. Every timed publisher now writes Info::default().with_timescale(Timescale::MILLI), and about 60 test modules carry a local timed() helper.
   • (a) Keep as is. Recommended: no consumer needs new API yet.
   • (b) Add a public track::Info::timed(scale) (or similar) and swap the helpers for it.
   • (c) Revisit in the publish-timestamp quest, which touches the same publishers.
2. Stale line in rs/moq-mux/AGENTS.md ("Info::default() is milliseconds"). The advice to pin with with_timescale still holds, but the reason is wrong now. Recommended: change it to "is untimed" in this PR once approved.
3. TS SI table timescale. These tables keep milliseconds, which is today's behavior, while their PTS is finer. Recommended: keep milliseconds here and settle it with the shared-clock quest.

Issues found along the way: the TS SI tracks (rs/moq-mux/src/container/ts/si.rs) were a production path that relied on the old default, and only the moq-cli import/export integration tests caught it. Local moq-uring tests fail on RLIMIT_MEMLOCK (environment only).

(Written by Claude Opus 5.5)