fix(runner): guard the test_terminal.eigs invocation (fixes #656)#659
Conversation
…s#656) The [60] Terminal Builtins invocation was the last unguarded interpreter call in the runner: it lacked both the $EIGS_TMO runaway backstop (InauguralSystems#616) and </dev/null. Because test_terminal.eigs reads stdin, any environment where stdin is a held-open pipe (a captured or backgrounded CI-ish shell) made it block forever, wedging the whole suite rather than failing. Give :2073 the same guarded shape as the check_eigs_suite call at :93 — $EIGS_TMO and </dev/null — and delete the now-false CLAUDE.md stdin rule, which documented this bug as an invocation-discipline constraint. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
Verified and merging. Fast turnaround — thank you. I checked the one claim that had to be true before deleting a documented rule, since removing the CLAUDE.md line means the runner has to be robust on its own. Your wording — "the last unguarded invocation shape in the runner" — is a bit loose: there are plenty of bare I confirmed it rather than reasoned about it, with a FIFO holding the write end open so the reader could be killed independently:
Your Worth noting what this issue was: a bug we had documented as a constraint and lived with. You paid 11 minutes to rediscover it because we wrote the workaround into CLAUDE.md instead of fixing the one line. Good find, and the coat removal is the better half of the PR. If you find the project useful or enjoyed contributing, a star helps others discover it. |
…640, #641) (#662) Each returned a plausible-but-wrong number at rc=0 — the highest-yield failure class. All four were found in the 2026-07 adversarial review and each fix ships with a regression test that fails on the pre-fix library. #639 punnett_square: Aa x Aa read 1:1:1:1 instead of the canonical 1:2:1. The two heterozygous boxes "Aa" and "aA" are the same genotype but were keyed distinctly. Canonicalize each genotype (dominant/lower-ord allele first) so they fold; Aa is now 0.5 with no separate "aA" key. #640 power_iteration: returned |lambda|, not lambda. It read the eigenvalue off |w| = |A v|, which drops the sign — diag(-5, 1) reported 5, which is not an eigenvalue of the matrix at all. Use the Rayleigh quotient (lambda = vᵀ(A v) = v·w for unit-norm v), summed from the pre-update v. Positive spectra are unaffected (both agree there). #638 correlation: with mismatched lengths it paired min(len) elements but centered them on the FULL-list means, so unpaired tail elements corrupted the covariance. Center on the paired prefix instead. #641 two degenerate-input contracts returning plausible-but-wrong numbers: - factorial of a negative returned 1 (via `n <= 1`); it is undefined, so return the 0 sentinel — impossible for a real factorial (always >= 1), matching combinations' return-0-on-invalid convention. - time_of_flight could return a NEGATIVE time when both roots of x = v0 t + ½ a t² are negative (target never reached going forward). Return the smallest NON-NEGATIVE root, or the 0 sentinel the function already uses for the no-real-root (disc < 0) case. Tests: 8 known-answer regressions added to test_stem_accuracy.eigs (each verified to fail with its fix reverted), plus test_biology's Punnett check strengthened to lock in 1:2:1 and assert no "aA" key. The stem-accuracy block emits no "Tests: N" line so its runner tally is a hand-synced literal (#654's documented-fallback class): bumped 123 -> 131, and guarded its invocation with $EIGS_TMO + </dev/null to match #659. Suite 3075/3075 release. No C changed; behavior is pure-.eigs library logic.
2 participants
Fixes #656 — thanks for first refusal. One line plus the coat removal:
matching the guarded
:93shape, and the CLAUDE.md stdin-folklore rule is deleted verbatim since it stops being true.Your four "done when" items, all demonstrated
[60] Terminal Builtins: 10/10 in a normal run.sleep 300 | bash run_all_tests.shunder a bounded outer timeout — BEFORE: did not complete, log frozen exactly at[60] Terminal Builtins, zero RESULTS lines, and zero orphaned interpreters after group-kill cleanup (pgrep -a eigenscriptempty). AFTER: the same held-open-stdin invocation runs to a greenRESULTS: 2891/2891. (Measurement note: with your exact repro the outer wrapper can't exit beforesleep 300does regardless of the fix — pipeline semantics — so "completes" is correctly judged by the suite emitting its RESULTS line.)[60]; restoring completes green — the guard is load-bearing, not coincidental.bash -nclean; shellcheck absent on this box (stated, per house custom by now). One test-harness pitfall worth passing on:setsidwithout--waitforks-and-exits so$!tracks a dead pid — the first BEFORE attempt falsely looked "completed" with an empty log;setsid --wait+ writing the inner PGID to a file for group-kill cleanup is the reliable shape.🤖 This PR was written with AI assistance (Claude), directed and verified by a human-supervised workflow.