Flaky: learn tournament E2E fails on a missing mutator-attempts-1 barrier file (same commit passes on re-run)

[justrach]

Observed on PR #465, 2026-08-06. Two zig CI runs against the identical head SHA (0fea4f3): one failed, one passed.

tournament E2E failed; fixture preserved at /tmp/codegraff-learn-tournament.ob35k_n2
    assert (mutation_failure_barrier / "mutator-attempts-1").read_text() == "2"
FileNotFoundError: [Errno 2] No such file or directory:
  '/tmp/codegraff-learn-tournament.<tmp>/barrier-mutation-failure/mutator-attempts-1'

Failing run: https://github.com/justrach/codegraff/actions/runs/31104029504/job/92624467497
Passing run, same SHA: https://github.com/justrach/codegraff/actions/runs/31104035680

Same commit, opposite outcomes, so this is genuinely non-deterministic rather than a content regression. The PR that surfaced it (#418) changed only a new test file, test_hooks.zig and comments, none of which the tournament path touches.

What the assertion wants: the barrier fixture expects the mutator to have been attempted twice, recorded as a counter file. The file is absent entirely, not holding a wrong value — so the failure is upstream of the count: either the barrier directory was never populated, or the assertion ran before the writer finished.

Given #444 established that this repo's other long-standing "flake" was a misread of build-runner output rather than real non-determinism, it is worth resisting the same assumption here. This one has hard evidence of true non-determinism (identical SHA, both outcomes), so the mechanism is real and worth finding rather than papering over with a retry.

Likely candidates, in the order worth checking:

1. A write/read race — the harness asserts on a counter file the mutator subprocess writes, without waiting for it. A missing file rather than a stale value fits this best.
2. Fixture teardown or /tmp reuse — the fixture path is preserved on failure, so a preserved-but-empty directory from the failing run would distinguish this quickly.
3. A spawn that failed silently, leaving the barrier unpopulated, with the assertion reporting the symptom rather than the cause.

Whoever takes this: the preserved fixture path in the failure output is the first thing to look at, and the two runs above give a working pair to diff.

[justrach]

Root cause confirmed as the write/read race option: the fixture mutator wrote its attempts counter only AFTER the four-way rendezvous loop — a peer lagging past the hardcoded 10s deadline died at its own assert before writing, so the E2E asserted on an absent file.\n\nFixed by 1fc6ecf: the attempt is recorded BEFORE the rendezvous and the deadline is env-tunable (TOURNAMENT_BARRIER_DEADLINE, default 120s). New deterministic regression tests (tests/test_learn_tournament_fixtures.py) force the lag via a tiny deadline and require the counter to exist regardless of rendezvous outcome.