fix(flapping-plists): sweep 4 flapping launchd plists — TCC wrapper drop + dashboard-dependency inline + exit-1 misclassification

[mitwilli-create]

Summary

Closes the R3 residual from the 2026-05-29 chain handover: 4 of 6 flapping launchd plists fixed in one same-PR sweep. Two remaining (buttons-smoke, scan-email-poll) need investigation; not addressed here.

Three failure modes, four fixes:

Plist	Pre-fix exit	Root cause	Fix
network-database-build	126 (since 2026-05-23, ~7 days silent flapping)	macOS Tahoe TCC blocks /bin/bash from reading scripts under ~/Documents/	Drop bash + cron-run.sh wrapper; invoke node directly. Matches bug-intake-mapper pattern.
phase-B-prime-daily	1 (since 2026-05-25, 5 consecutive days)	_CONTACTS_DATA not populated when 03:30 PT job fires before morning rebuild	Inline node scripts/build-dashboard.mjs at start of main(). Override via SKIP_DASHBOARD_PRELOAD=1.
pipeline-health	1 (false-positive flap when 2+ concurrent sessions)	Hardcoded pids.length > 1 orchestrator threshold	Env-configurable PIPELINE_HEALTH_MAX_ORCHESTRATORS (default 1) + process.exit(0) always. The JSON file IS the signal per top-of-file comment.
health-column-liveness	1 (false-positive flap when coverage <90%)	Exit 1 for valid data-coverage signal misclassified as crash	process.exit(0) on unhealthy. Keep exit 2 for true FATAL. JSON at data/health-column-coverage.json is the signal.

AGENTS.md additions

Two new bug-class entries:

• ### Bug class: launchd-bash-wrapper-tahoe-tcc-block — full doc of why bash hits TCC and node doesn't, with safe-pattern XML.
• ### Bug class: launchd-exit-1-misclassified-as-flapping-on-data-signals — generalizable pattern: when a script writes its real signal to a JSON file AND returns an exit code, treat the file as the signal. Reserve non-zero for true script failure. Companion pattern: env-configurable thresholds replace hardcoded "expected 0 or 1" assertions.

Smoke tests (all pass)

plutil -lint:           4/4 plists OK
node --check:           3/3 .mjs files OK
pipeline-health exit:   0 (was 1; JSON signal preserved)
health-column exit:     0 (was 1; JSON signal preserved)
PIPELINE_HEALTH_MAX_ORCHESTRATORS env knob: honored

Test plan

• Merge the PR
• Manually trigger each fixed plist to confirm flapping resolved:
  • launchctl kickstart -k gui/$(id -u)/com.mitchell.career-ops.network-database-build → exit 0 expected
  • launchctl kickstart -k gui/$(id -u)/com.mitchell.career-ops.phase-B-prime-daily → exit 0 expected (cost ~$2.50 if run during work day)
  • launchctl kickstart -k gui/$(id -u)/com.mitchell.career-ops.pipeline-health → exit 0 expected
  • launchctl kickstart -k gui/$(id -u)/com.mitchell.career-ops.health-column-liveness → exit 0 expected
• After 24h, re-run node scripts/agents/system-maintainer.mjs --health → should show 2 flapping instead of 6
• (Optional) Set launchctl setenv PIPELINE_HEALTH_MAX_ORCHESTRATORS 2 if concurrent-session orchestrators are routine

Out of scope (follow-ups)

• buttons-smoke: 1/14 assertion failure ("batch-runner dry-run reports expected queue size") — code investigation, not a crash. File a separate PR.
• scan-email-poll: produces no .err / .out logs at all — likely plist misconfiguration or job stub that never starts. Investigation needed.

🤖 Generated with Claude Code