[kixelated]

Finishes quest/m0/queue-watcher/flaky-signature.md.

Summary

The watcher matched a flaky verdict to its test's record by signature(failure), so a fixer that reworded a known test (with or without the command that ran it) started a new record, and no test reached FLAKY_VERDICTS.

• A flaky requeue of a known test now carries that known entry's quest (quest/m0/flaky/<name>.md), copied from the input instead of retyping the name. recordFlaky finds the record by quest; only a requeue with no quest falls back to the signature.
• parseFixResult takes the known list and rejects a quest that is not a known one. disable now requires quest, since only a listed test may be disabled. A flaky requeue must name its test (failure or quest), so it never falls back to the job's step name and merges unrelated tests.
• assets/land-fix.md: name a new failure by its package and test alone, without the command, and add quest for a known test.
• Test: the disable test's four verdicts use the wordings seen on moq-dev/moq (cargo nextest run --locked --tests: render_timeout, moq-relay render_timeout, render_timeout) and land in one record that disables at the fourth.

Public API and wire impact

The fixer's result format gains an optional quest field on a flaky requeue; disable without it is now rejected. Internal to quest land, nothing in production.

Choices made without asking

• Dropped nothing: the first flaky verdict guard in recordFlaky stays, though a known test always has a verdict already, because a re-collected job replaces its own verdict.
• Kept the signature fallback for a new test rather than stripping a command prefix heuristically: the prompt change asks for the test alone, and a parser for every runner's prefix would be guesswork.

Checks

bun run check (tsc, prettier), quest check, and vitest run test/land.test.ts test/land-watch.test.ts (247 passed) ran locally. The full just check / just test sat in quest gate behind host load (52) for over an hour and did not run; CI covers them.

Decisions

• Cause of the split records: ✅ accept the code reading (rejected: confirm from flaky.json on the watcher's host first). A record is created only on a signature miss, so rewording a known test is the only way records split. Nobody checked flaky.json on the watcher's host.

Models

• Implementer: Claude Code, claude-opus-5-5 (iterate, S)

(Written by Claude Opus 5.5)

[pullfrog]

  ^{｜ Review this ➔}