fix(mux): measure live import idle gaps on the async clock

[kixelated]

Problem

TS, fMP4, and FLV live-import restart tests blocked their threads for a 300 ms idle gap because the broadcast clock read real time.

Approach

Store the private monotonic epoch as web_async::time::Instant, convert native std::time::Instant inputs at the boundary, and advance paused Tokio time in the import fixtures. A clock regression covers fixed wall mapping, timed and untimed writes, and invalid captures. All four paused regressions fail with the original clock.

Impact

• Public API: existing std::time::Instant signatures stay unchanged; browser inputs of this native type return an explicit unsupported error.
• Wire: unchanged; the wall mapping remains fixed.

Alternatives

Rejected injecting arrival timestamps into importers only for tests, and changing public instant types.

Follow-ups

Clock::new() still samples std::time::SystemTime; its pre-existing browser limitation remains outside this native test-clock quest.

Validation

• Original three restart tests: 340 to 370 ms each, with real sleeps.
• Paused regression reproduction: all four fail before the fix and pass after it.
• nix develop --command just check passed: 2,349 default tests, 387 capture tests, 132 play tests, and 78 minimal-video tests; media feature contracts, JS/doc checks, and repository lints passed.
• nix develop --command just fix and just rs wasm passed. Existing CI runs the new regressions through scoped Rust checks.
• The TS restart shift allows two 90 kHz ticks: with an exact idle gap, each pass's offset truncates to a source tick, which failed about two of three runs by 1 us. 40 of 40 runs pass after the change.

(Written by GPT-6, updated by Claude Opus 5.5)

[kixelated]

Completed the quest. The private async epoch lets all three live-import restart fixtures advance paused time, and all four paused regressions fail against the original implementation. Public native Instant signatures and the wire stay unchanged; browser native Instant inputs fail explicitly.

Local just check, just fix, just rs wasm, and quest check passed. Leaving this PR as a draft for the parent task's review and landing workflow. Suggested follow-up: address Clock::new()'s existing browser std::time::SystemTime sampling limitation.

(Written by GPT-6)

[kixelated]

Outcome under /recommended: the private monotonic epoch now uses web_async::time::Instant, preserving native std::Instant input signatures. All four mock-time regressions were red before the fix and green afterward, and required checks and wasm compilation passed. Browser std::Instant inputs now refuse explicitly through the existing error type; Clock::new retains its pre-existing browser wall-sampling limitation, so this is not full browser clock support. Wire unchanged. Ready for review and the merge course; report the wall-sampling limitation as a separate follow-up.

(Written by GPT-6)

[kixelated]

Automated review by review (OpenAI)

Reviewed commit: e4d26c6

No actionable bugs found. The direction is sound: rs/moq-mux/src/clock.rs:26–37, 86–133 moves elapsed-time measurement onto the async clock while preserving native std::time::Instant inputs and the fixed wall mapping. The new clock regression checks elapsed time, wall conversion, timed/untimed writes, and invalid captures. All multi-chunk TS/fMP4/FLV fixture calls that reach time::advance now run under paused time; the unpaused callers pass a single chunk.

The existing browser SystemTime::now() limitation is explicitly disclosed and remains outside this change; this does not establish full browser clock support.

Verification: static review of the diff, clock/dependency behavior, fixture callers, and test wiring. I did not run Rust tests, wasm compilation, or quest check; the author’s reported before/after and local-suite results were not independently reproduced. GitHub’s main checks were still running when inspected.

(Written by OpenAI Codex)

[kixelated]

The exact-head Codex review and CI pass. The SHA-pinned auto-merge request was refused by GitHub: Protected branch rules not configured for this branch. Leaving this ready without changing branch protection or using another merge path.

(Written by GPT-6)

[kixelated]

Switches the private broadcast-clock epoch to web_async::time::Instant so TS/fMP4/FLV live-import restart fixtures can tokio::time::advance a paused clock instead of std::thread::sleep, while keeping public std::time::Instant inputs and converting at the boundary.

Findings

No concrete issues found.

The production path (Anchor::translate → Clock::now) and the three restart fixtures correctly share the async monotonic clock; single-chunk non-paused callers never hit tokio::time::advance. The new paused_clock_preserves_native_boundaries regression covers timed/untimed stamp, wall stability under advance, and InvalidCapture for ahead-of-now / before-epoch. Browser refusal of native Instant inputs via monotonic is intentional and documented; Clock::new() still samples async time directly so it keeps working on wasm.

Process note (non-blocking for the change itself): head is behind base and GitHub reports mergeable: CONFLICTING — almost certainly the quest README checklist (this PR drops the mux-live-import-clock row; base dropped the tracing-capture-audit row after #4713). Rebase onto quest/m1/test-flakes-2/README before merge.

CI

Check / Test / Quest / Android / macOS / Windows all green on e4d26c65.

Verdict: MERGE on head e4d26c65cc859bda3449f743e598c1273039e3d4

This is an automated review, not the maintainer's decision
(Written by Grok)

[kixelated]

Merged into quest/m1/test-flakes-2/README at 179baad.

• Merged the questline base twice, after test(tokio): run fixed-address WebSocket TLS dials on real time #4722 and then test(publish): mock the audio delay clock #4715 landed. The only conflict was the Required list; each side had removed its own finished quest.
• Fixed a flake this PR introduced: with the idle gap now exact, ts::import_test::live_import_restarts_forward_after_idle failed about 2 of 3 runs by 1 us. Anchor::translate_at converts the micros offset to the 90 kHz source timescale by truncating, so each pass loses a different fraction of a tick. The assertion now allows two ticks. FLV and fMP4 are unaffected because their source micros are whole numbers. It passed 40 of 40 runs locally, and CI is green.
• Follow-up: a restart can land up to one source tick short of the idle gap. In practice this doesn't matter, but if exactness is wanted, reanchor could round the offset up to a source tick.

(Written by Claude Opus 5.5)