test(timing): measure CPU time and corroborate before failing on loaded runners by rennf93 · Pull Request #163 · Guard-Core/guard-core

rennf93 · 2026-10-15T01:38:24Z

Master CI failed its last two runs (2026-10-14) on timing assertions that a loaded runner can miss, and PR #162 hit the same one. The failures, verbatim from runs 38613193702, 38593971238 and 38639913718:

test_lazy_init_returns_quickly_with_blocking_background_init - assert 0.6880667630000517 < 0.5 (also 0.5635 and 0.6231)
test_bounded_scan_stays_under_five_milliseconds_on_all_three_fill_shapes - many_prefixes_one_terminator at size=524288: bounded CPU 0.005967s exceeded the 0.005s ceiling
docker-test: hostname fragment timed out with failing tail, and the template and ldap extensible-match detect did not return within the deadline tests

Changes, each keeping the test able to fail on a real regression:

Lazy init: the stub never blocked, so the 0.5s bound could not catch a regression either. The background init now blocks for 2s and the test asserts ordering (the call returns before the background init finishes). A lazy init that awaited it fails with assert ['cloud'] == []; the old test passed under that mutation.
Bounded scan: the 5ms ceiling scales linearly with size and must be breached at every size, so a single contention-inflated min-of-5 no longer fails it. A 4x slowdown of bounded_search fails at both sizes, and swapping in a raw compiled.search fails as well. _REPS is unchanged and there is no retry.
ReDoS protection and hostname fragment: the deadline is now the child's CPU time around detect() (time.process_time()), and wall clock only guards against a hang (60s), since forkserver start-up and imports alone outran the 1-2s wall join under Docker load. An injected (a+)+$ scan fails with detect used 4.459s of CPU, exceeding the 2.0s deadline, and the nested-star hostname probe fails with used 3.776s of CPU.

The sync tree is regenerated with scripts/unasync.py (the hand-maintained ReDoS mirror edited by hand), and --check passes. Local gate: 21568 passed, 4 skipped, 100% coverage, quality and pre-commit clean.

…ed runners The lazy-init returns-quickly test asserted a 0.5s wall bound that a loaded runner exceeded (0.56 to 0.69s) while its stub never blocked, so it could not fail on a real regression either. It now blocks the background init for 2s and asserts ordering: the call returns while that init is unfinished. The scan-window 5ms CPU ceiling is scaled per size and must be breached at every size, so one contention-inflated min-of-5 no longer fails it while a quadratic or constant-factor regression still does. The built-in ReDoS detect tests and the hostname fragment test joined a forkserver child under a 1s or 2s wall deadline, which start-up alone outruns under load. The wall join is now a 60s hang guard and the budget is the child's CPU time.

github-actions · 2026-10-15T02:27:29Z

Detection gate report

All detection gate shards passed.

Shard `arbiter-grammar`

Gate green: no category regressed against its committed baseline and no benign string is newly flagged.

Total recall: n/a. Total false-positive rate: n/a.
Documented known gaps: 0. Documented known false positives: 0. Both stay in the denominators.