[Mening-dotcom]

Problem / Motivation

Goal: The idle-seed tests stop failing a Windows shard when a scanner briefly holds a freshly written temp file.

test_private_transcript_is_not_seeded[temporary] failed Backend Tests (Windows) (2) with PermissionError: [WinError 5] from atomic_write.py's replace, on a PR that changes label text only (#24915; run 40569051980). The traceback runs test → _restarted → update_metadata → _update_metadata_locked → the replace. The cause is the test setup's position, not the writer: replace_with_retry deliberately skips its bounded Windows retry when a loop is running in the thread (re-raising instead of sleeping on the gateway loop), and _restarted performs its transcript writes synchronously inside an async test under allow_on_loop_persist() — on the loop. One transient AV/indexer handle on the fresh temp therefore fails the test on the first attempt, where the same write off the loop would have retried through the 10 × 50 ms budget and passed. (#24977)

Why it matters

The shard was red on a PR that touched nothing near history consolidation — exactly the kind of unrelated-red that trains people to re-run CI instead of reading it, on the OS where rename contention is routine.

Not a goal

• No change to replace_with_retry or its on-loop re-raise: that behavior is the repo's no-blocking-call-on-event-loop anchor working as documented, and the documented remedy is for the caller to offload.
• No change to the file's other three allow_on_loop_persist blocks: _write_row's single append and the one-row delete are each the behavior under test (same-process writer semantics, the dashboard delete), and the no-op mark_consolidated(key, 0) is asserted byte-identical — none is a fixture bulk-write, and none is the frame the flake fired in.

What changed (motivation → approach → change)

Observed: the two fixture writers (_restarted, _live) persist transcripts on the event loop, where the Windows replace retry cannot apply. Root cause: the setup stands for "an earlier process wrote these transcripts", and an earlier process has no event loop — the on-loop position is an artifact of calling the sync helper from an async test. Change: each helper's write block now runs in a bounded worker (ThreadPoolExecutor, result(timeout=30) so exceptions propagate and a hang fails rather than deadlocks). Off the loop, the writes earn replace_with_retry's bounded retry, and the allow_on_loop_persist() hatch becomes unnecessary for these blocks and is dropped there (it remains where the on-loop position is the point).

Backwards compatibility

Test-only; fixture semantics unchanged (same records, same files — written by a worker thread exactly as the "earlier process" they model would have). All 29 tests in the file pass on native Windows; the flake-pattern ratchet passes (test_flake_pattern_ratchet, 62 passed).

Tests

• test/test_history_consolidation_idle_seed.py on native Windows (the failing platform): 29 passed.
• python -m pytest -n0 test/test_flake_pattern_ratchet.py: 62 passed.
• scripts/check_comment_history.py, black --check, isort --check-only clean.

Manual verification

N/A — test-only change, verified by running the file on the platform the flake reports.

Screenshots / video

N/A — no user-facing surface.

Related Issues

Fixes #24977

Pattern harvest

Rule candidate: review-prompt
Pattern: a sync fixture helper called from an async test runs its file writes ON the event loop, where replace_with_retry deliberately re-raises instead of retrying — any allow_on_loop_persist() around fixture bulk-writes is a sign the setup should be a bounded worker-thread call instead.

Checklist

• At most two commits (one is the norm), with a Conventional Commits title (feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)
• Existing tests pass and new tests added for new functionality
• Self-review completed; code follows project style guidelines
• Documentation updated (if applicable)
• No secrets, credentials, or internal references in the diff

Contribution License Agreement

Contributions accepted under the project's Apache-2.0 license per CONTRIBUTING.md.

[Mening-dotcom]

Closing: current main already settles this — #23788 moved the idle-seed fixture writes into an _off_loop worker helper (same diagnosis: on the loop, replace_with_retry re-raises instead of retrying the Windows rename), so this branch now duplicates it. Note for #24977 triage: the fix is on main via #23788 and the issue can likely be closed.