[matt2e]

Summary

Fixes two Rust test flakes from the 23–30 Oct CI flake sweep. The first is a real defect in the media spool, so this PR changes product code. The second is test-only. This is one of six independent PRs split from the flakey-tests branch. Each is cut from main and can merge in any order.

A failed spool still answers the bytes it wrote (media_stream_tests.rs:277)

• CI: 39970306754 (job 120470903732), 40829345046 (job 123299699319), 40958827040 (job 123745997146). a_failure_mid_stream_ends_the_response_short_of_its_length fails at :300 with 502 where 200 is expected. The case is a 200 response that sends 2 MiB of a 3 MiB Content-Length.
• Cause: the spool reports progress through a tokio watch channel, which keeps only the latest value. The copy flushes the first MiB as Writing(n), then fails and overwrites that with Failed(502). A reader that polls only after both sends sees just Failed. read() then returned 502, even though the bytes it asked for were already on disk, and the response's first block went out as a 502 head. A player's range request can hit the same race in the app.
• Change (media_spool.rs): Failed also carries the last byte count the copy reported. read() answers from disk whenever that count covers its range, and otherwise returns the failure status as before. lookup() still drops a failed spool, so later reads fetch the video again. A read past the written bytes still fails, and truncated_whole_videos_fail_and_are_fetched_again is unchanged.

The memory-read deadline test no longer depends on decode speed (agent.rs)

• CI: 40877681973 (job 123460621816) got 16 entries where 32 were expected. It passed on re-run.
• Cause: each page took 120 ms against a 300 ms deadline, and page 1 is decoded before page 2 is requested. Page 2 therefore made the deadline only if page 1's decode finished in about 60 ms, and a slow decode on CI pushed it past.
• Change: read_through answers each page at once and leaves the request after the last page pending. The first case now has 32 events, two full pages, so the deadline always passes while the third page is awaited. The second case asserts the same as before.
• Hang guard: because the last request now stays pending, a read that stopped honouring its deadline would hang instead of failing its took bound, and CI runs plain cargo test with no per-test timeout. read_through therefore wraps the read in a 5 s tokio::time::timeout that fails with "memory read outlived its deadline". The took assertions still check the real bound.

Verification (local macOS)

• The new test a_failed_spool_still_answers_the_bytes_it_wrote forces the ordering. It first reads the tail, which can only return after the copy has failed, and then reads the first block, so the second read never sees Writing. Without the media_spool.rs change it fails on the first-block read with called `Result::unwrap()` on an `Err` value: 502. With the change it passes. :277 and truncated_whole_videos_fail_and_are_fetched_again also pass.
• Memory read: a temporary probe (not committed) added a 100 ms decode stall after page 1. The old test then failed with left: 16, right: 32, the CI signature, and the new test passed. A second probe moved the deadline 1000× later. The test failed after 5.05 s with the hang-guard message instead of hanging.
• Ran cargo fmt --check and cargo clippy --all-targets -D warnings on the Buzz crate, plus the four tests above by name. After the hang guard was added, only cargo fmt --check and the memory-read test were re-run. Clippy and the media tests were not.

Overlaps

Drafts #1461 and #967 rework relay_agent_memories_read, and #1461 also edits media_stream.rs. Neither changes the code paths fixed here.

Remaining gaps

• The full Rust suite has not run locally. Hosted CI is its first full run.
• A green run does not show that these flakes are gone.

🤖 Generated with Claude Code

[wesbillman]

🤖 Star Lord’s automated source review via Wes’s account, integrating independent Rocket (systems/lifecycle), Mantis (correctness/security/concurrency), and Groot (product/error paths) reviews.

Reviewed head 1927a0208dd4501b5319ddbb8a16b88fd1d17964 against base 36a283b2386473130430890faa3b06c6e45d4a6b. No actionable findings; no changes requested. The failure state preserves the last completed disk-write boundary, allowing an existing reader to answer a fully written range while missing bytes still fail. Failed spools remain unavailable to new cache lookups. The added regression forces failure before reading the available prefix; the memory test retains its elapsed-time assertions while replacing delayed page delivery with a pending final page and a bounded hang guard. No browser cases or test infrastructure were changed.

Source-only review of the three-file diff, callers, cancellation/file lifetime, network-error recovery, repository guidance, and public description/commit text (no attached images). Hosted CI snapshot: CI required succeeded; Rust job 124048893021 logs show the four relevant tests passed and the Buzz crate reported 702 passed, 2 ignored. Windows native validation and advisory flake lanes were skipped. We ran no local tests or PR code and did not launch the app. Native playback, error/retry focus behavior, cross-platform acceptance, and long-term flake elimination remain unverified; the memory test still uses real-clock deadlines. This COMMENT is not approval or merge authorization.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

Reviewed head ec3a326ae8d7da9c0d37cdd46dcc162c0a505c4f against base f4805605f56a2f60bd25b966464f07f0c3292a41. No new actionable findings; no changes requested. The spool, block reader and loopback stream implementation are unchanged from the earlier reviewed revision. I checked the test’s integration with the rebased pipelined memory reader and its larger pages, along with media failure/retry consumers. Optional description clarification: the current first memory-test case uses 256 events (two 128-event pages); the historical 16/32 failure figures can stay identified as historical.

Hosted evidence at this head: the four relevant tests passed, and the Buzz crate reported 848 passed, 2 ignored in 38.12s. The Rust job nevertheless failed in crates/agent-command/tests/discovery.rs:438 with ExecutableFileBusy (outside this patch). Browser jobs were still running in the status snapshot; Windows validation was skipped. No green-CI claim.

Source-only review; no local tests, builds or app execution. Public description and feature commit text inspected; no attached images. Native playback, error/retry focus, cross-platform behavior and lasting flake elimination remain unverified. The memory test still uses real-clock deadlines; hosted success does not prove deterministic timing. This COMMENT is not approval or merge authorization.

[wesbillman]

🤖 Star Lord’s automated source follow-up via Wes’s account, integrating independent Rocket, Mantis and Groot review lanes.

Reviewed head 19f79ebca0e9b18f4c116c661427bdd2dda95de7 against base 56729773f48d3f922a043ac83b7faa727bcfc9c2. No new actionable findings; no changes requested. The native relay subtree, media entry point and playback/recovery consumers are unchanged from reviewed head ec3a326ae8d7da9c0d37cdd46dcc162c0a505c4f. This follow-up checked the rebased integration; it does not reopen unrelated main-branch changes. The failed-spool byte boundary, retry/cache eviction and memory deadline-test behavior remain intact in source.

Source-only: no local tests, builds or app execution. Public description and feature commit text inspected; no attached media. Native playback, error/retry focus, performance, cross-platform behavior and lasting flake elimination remain unverified. The current-head CI snapshot still had Rust, JavaScript and browser jobs running; Windows and advisory flake lanes were skipped. Final test counts and timings were unavailable at that snapshot. This COMMENT is not approval or merge authorization.