Harden guard concurrency diagnostics

[screenleon]

Summary

• add bounded child tracking, per-case deadlines, TERM-to-KILL cleanup, and actionable PID diagnostics to test-guards.sh
• migrate every background writer/FIFO case away from unbounded wait
• add deterministic regressions for a hung writer and a nonzero child exit
• enable per-case CI breadcrumbs and cap the direct test-guards job at 15 minutes

Root cause

The direct GitHub Actions guard job exposed only a section heading while several concurrency cases used background writers, FIFO coordination, and bare wait. If a writer failed to converge, the job had neither a case-level deadline nor enough process context to identify and reap the child, allowing one case to consume nearly the entire runner window.

Impact

Guard concurrency failures now terminate within a bounded deadline and report the active case, backend, round, writer label, PID, process state, lock state, and completion counts. Normal correctness assertions, writer counts, flock coverage, and mkdir-fallback coverage remain unchanged.

Validation

• focused hang and exit-7 fault-injection regressions: PASS
• contention matrix stress: 3 repeated runs PASS
• affected runner: 2/2 suites PASS (test-guards 300/300)
• standard PR gate gate-20260714-075600-a3affd: Final GO
• targeted remediation gate gate-20260714-080619-5069e6: Final GO
• authoritative full suite: 77/77 suites PASS, 0 failed, 0 skipped
• final tree artifact verified at fingerprint 79a61fbb6b1566c2e80dca00d307346bc727e0ce3c510d72cbee447fa71b7f78

One intermediate full run observed an unrelated load-sensitive test-run-all-tests failure; that suite passed standalone 34/34, and the unchanged tree then completed a clean authoritative 77/77 full run.