fix: idle-timeout for agent-loop engine calls, infra-failure escalation

[Santisoutoo]

Summary

• run-engine.sh: replace the flat 15min wall-clock timeout (PR fix: bound engine calls with a per-call timeout in run-engine.sh #102) with an idle-timeout — a call is only killed after N seconds with zero new stdout/stderr, with a much larger hard ceiling as backstop. Fixes today's incident where the Implement step was killed mid-exploration on issue [data-fidelity] Sprint 1 — metadata trust: status enum, 4KB frontmatter, article identity #52 while still actively producing output.
• Per-step budgets: Implement idle=6min/ceiling=40min (needs exploration room), Review idle=3min/ceiling=15min (this is where the real 84min zero-output hang happened, issue [frontend-perf] Sprint 4 — bundle hygiene: mock out of entry, lazy wizard, drop dead dep #85), fix-round idle=5min/ceiling=20min.
• Fix a misclassification: a DONE-but-empty-diff guard failure (known opencode silent-exit bug, anomalyco/opencode#28605) was wrongly counted as a real attempt instead of the infra failure it is.
• Infra failures (engine crash, timeout, no result marker) now get their own counter and escalate to a new agent:infra-stuck label after 3 — previously they could retry forever, 3x/day, with no escalation.
• Every failure path now writes a short reason into the issue comment (implement-idle-timeout, verify-backend, review-fix-exhausted, ...) instead of a generic message — no more opening the Actions log to find out which stage broke.
• Verify — landing now installs Chromium before building, matching ci.yml's landing-build job.
• New Verify — pip-audit (CVE gate) step: confirmed via gh api .../branches/main/protection that pip-audit isn't a required branch-protection context, and agent-loop never checked it either — a CVE-flagged dependency could previously auto-merge unblocked. Verified main is currently clean (uv run --with pip-audit pip-audit --skip-editable → "No known vulnerabilities found") before wiring this in as blocking.
• pick-issue.sh: excludes the new agent:infra-stuck label from future picks. README.md and scripts/setup-github.sh updated to match (label already created on the repo).

Validation done locally

• bash -n on all three shell scripts, YAML-parsed the workflow (js-yaml) — all clean.
• Standalone smoke test of the idle-detection loop shape: a simulated hang (output stops, then 60s of silence) gets killed after the idle threshold (~7s in test config), not the full hang duration. A simulated slow-but-steady process (output every 2s) runs to completion untouched. Confirms the core fix for today's false-positive.
• pip-audit run locally against main: no known vulnerabilities.

Known residual risk (documented, not blocking)

run-engine.sh's idle-kill uses pkill -TERM -P "$runner_pid", which only reaps direct children, not grandchildren — if opencode/cursor-agent spawn multi-level subprocess trees, an idle-kill could theoretically leave an orphan running. Plan is to validate via a workflow_dispatch test run before trusting the 3x/day cron on this, per the README's own existing practice for engine/model changes, and watch the job log for orphan processes.

Test plan

• workflow_dispatch on a low-risk issue to confirm: idle-timeout doesn't kill active work, hard ceiling still works as backstop, no orphan processes, failure-reason appears correctly in issue comments if a failure is forced.
• Monitor the next scheduled run (3x/day) and confirm issue [data-fidelity] Sprint 1 — metadata trust: status enum, 4KB frontmatter, article identity #52 (or whatever's queued) is no longer killed mid-exploration.

Claude-Session: https://claude.ai/code/session_01KpuHkuMadFm8SGgw7nuZki