ci: speed validation and fix daemon exit detection
Summary
Use 8-vCPU runners for Coverage and run numeric ranges independently. Also fix the daemon-exit observation race exposed by Integration Tests: a PID released during the process snapshot now proves exit on a final liveness check. Unreadable observations for live processes still fail closed. No retry or timeout changes.
Three files. Scope expanded into daemon-contracts and its regression tests to resolve the CI blocker. No linked issue.
Two successful runner/layout trials on identical sources before the retirement fix:
| Completion | Existing | Proposed | Median gain |
|---|---|---|---|
| Coverage | 364 / 385s | 200 / 208s | 46% |
| Guards | 277 / 279s | 187 / 221s | 27% |
Coverage, guard controls, split guards. Requested Coverage vCPU-seconds increase about 9%; 16 vCPU costs more without improving the optimized critical path.
Validation
Tested 80dc3c3dde9ca96deec2337ec3435430dc912aba:
pnpm check:affected --run --head 80dc3c3dde9ca96deec2337ec3435430dc912aba: 42 passed, 25 skipped; receiptrun-n7ETdM,exactCommit=true,checkoutChanged=false.- Same focused command against old/new implementations: 19 passed, 2 failed → 21 passed. Real-daemon cleanup integration: 3 passed. Independent review: no findings.
- Original failure reproduced on base and old head.
CI and all reported platform checks passed at this head. Cubic found no issues; Apex approved. Its minor PID-reuse finding was rejected with contract evidence and resolved.