run_all_tests.sh: .eigs test blocks have no runaway guard — an infinite loop STALLS the suite instead of failing it

[InauguralPhysicist]

Finding

The runner guards the examples against a runaway but not the test blocks. An infinite loop in any .eigs test file hangs the entire suite until the CI job's hard limit, instead of reporting a failure.

# tests/run_all_tests.sh:79 — check_eigs_suite, used by every .eigs test block:
out=$(./eigenscript "../tests/$file" </dev/null 2>&1); rc=$?     # <- no guard

# tests/run_all_tests.sh:3074 — the [97] examples section:
if command -v timeout >/dev/null 2>&1; then EX_TMO="timeout 60"  # <- guarded

Same runner, same risk, opposite treatment.

Why it's not hypothetical

#637 was exactly this bug in shipped code: pad_left with an empty pad char spun forever. Its regression tests (PL4/PR5, added in #646) sit inside check_eigs_suite. Verified: reverting lib/string.eigs while keeping those tests makes the suite hang — exit=124 only because I wrapped it in timeout by hand.

So today's protection against a #637 regression is "CI eventually times out and someone reads the log", not "a test fails." The difference is a red X versus a mystery — and the mystery costs a debugging session before anyone even learns which block stalled.

Why it matters more than it looks

Two of today's flakes (#616, #604) each cost a review cycle largely because a red result was ambiguous. A stall is worse than either: no failing assertion, no section name, no output — just a job that dies at the limit. It's the same failure class this repo already decided against everywhere else (rc_ok gates on exit codes precisely so a crash after correct output can't pass).

Fix

Wrap check_eigs_suite's invocation the way [97] already does — the pattern, the fallback (gtimeout on BSD userland), and the "no wrapper if absent" degradation are all already written 3000 lines below:

out=$($EIGS_TMO ./eigenscript "../tests/$file" </dev/null 2>&1); rc=$?

with rc=124 reported as a timeout failure naming the block, rather than folding into the generic path.

Pick the budget deliberately. #616 is open precisely because [97]'s timeout 60 is too tight for invariant_weak.eigs under ASan (60.5s standalone). A test-block guard is a runaway backstop, not a latency assertion — it should be generous (180s+) so it never fires on a slow-but-working test on the N3350. Worth fixing #616's guard in the same pass, since it's the same judgment.

Test

A scratch .eigs with loop while 1 == 1: run through check_eigs_suite must report a named failure and let the suite continue, rather than hang.

---

Found by @Nitjsefnie while fixing #637 (#646) — spotted as an out-of-scope note on their own PR, which is exactly where the good ones come from.