Repository navigation
oracle-gaps batch 3: default-out stop and daemon results missing beside the stream log (intermittent) #76
Description
Observed once, on PR #57 (feat/gap-cli-stream-log) at 10:50 on 2026-11-13, in a local batch (log batch3-r1.log, not in the repo).
Failing units and messages:
- fixed seed 20261105: a stop with the default output directory: "the run completed, but its results.json is not beside its stream log"
- fixed seed 20261105: a daemon check with the default output directory: the same message
- fixed seed 20261105: two runs on one output directory: "the second run on a live output directory exited 0, not refusing it"
Not reproduced in four later runs of the same code, two of them with two copies running at once.
Likely cause: a leaked mutant. Both failure signatures match the hostile mutants F2 (the supervisor does not pass --out to the child or daemon) and F4 (no live-run refusal). A bisect file swap that was in place at the time was ruled out by timestamps, but a mutant left in the file would produce exactly these messages.
Next step: make the results-beside-log message name the directory and whether results.json is missing or unparseable, so the next occurrence says which it was. Then find a check that asserts the file is restored after each mutant run.
Activity
fix(oracle-gaps): take over a stale claim under a mutex, and name res…
KyleAMathews commentedon Nov 14, 2026
Reproduced, and fixed by #57 (feat/gap-cli-stream-log, head 7db86c5).
Mechanism: the overlap witness kept the first run alive with a wall-clock hold (HOLD_EXIT_MS), so under load the second run could start after the first had exited. The CLI then correctly found a dead claim and took it over, and the witness's expected refusal was wrong for that schedule. That is a nondeterministic witness, not a CLI defect.
RED (parent 5fa74e0, unchanged CLI; second spawn delayed until the first process has exited, the extreme-load schedule):
cd <scratch worktree at 5fa74e0>/packages/oracles
npx vitest --run --config oracle-gaps/vitest.config.mjs gap-report-oracle -t 'fixed seed 20261105: two runs on one output directory$'
-> Difference: the second run on a live output directory exited 0, not refusing it
GREEN (7db86c5; the first run holds its exit and claim at a named point until the driver releases it after the second run is refused; same delayed second spawn, with an injected 3 s delay before it):
-> Tests 1 passed
Fix in #57: the overlap is deterministic by construction (ORACLE_GAPS_TEST_HOLD_UNTIL replaces the wall-clock HOLD_EXIT_MS). Status: fixed on #57; this issue closes when #57 merges.