fix(dor): un-shadow the died-sidekick check in the reconcile sweep

[TaekeK]

Closes #995

The 💀 liveness rule added in #963 has been unreachable for its entire target population since it shipped.

The bug

dor_reconcile.sh:143-150 flagged "no state:* label" as un-routed and continued — skipping the drift check, the whole case "$status" block including the "Building" liveness arm (:170), and the zombie-sidekick check (:183).

But "no state label" is not un-routed. The build side deliberately runs without one: dor_build_flow.sh drops state:awaiting-approval when it applies build-done, nothing re-adds a state label, and dor_feedback_flow.sh:32 sets board Status to Building on feedback. So an issue in exactly the state the rule describes — Status Building, sidekick dead — always arrived here label-less and could never reach the check written for it.

stalled is only populated inside that arm, so the dor-stuck label and the direct comment at :240-244 — the only signal that reaches a human — were dead for the same reason. Both pipelines are affected: dor-reconcile.yml runs the script for enhancement and again for bug.

Observed

On #370, runs 31016109943 and 31076341084 died when their sidekick was reclaimed. The sweep reported (health issue #886, 6 Aug 09:09):

🕳️ #370 is on the board (Status: Building) with no state:* label after 1151h — the agent likely never ran.

It had Status: Building in hand and still took the un-routed branch. The work sat dead ~18 hours across ~18 sweeps until @WimvandenHeijkant noticed by hand and re-posted both rounds of the requestor's feedback himself.

The fix

Ask the board whether anything owns the issue rather than trusting the label. A new build_phase() covers the four Statuses the build side owns (Building, Awaiting functional acceptance, Awaiting merge, Exceptions), and a label-less issue in one of them now falls through to the liveness and zombie checks.

• Genuinely un-routed issues still flag exactly as before.
• The drift check is now explicitly label-guarded (it was already a no-op without a label; made explicit so it stays that way).
• Un-routed age comes from updated_epoch, not created_epoch. That "1151h" was measured from June, for something that had been dead 18 hours — it read as ancient backlog rather than something that had just broken.

Tests

test/ci-scripts/test-dor-reconcile-liveness.sh — a first harness for .github/scripts/*.sh. The absence of one is why this shipped shadowed and stayed that way for weeks. It puts a stub gh on PATH serving fixtures, drives the real script end to end, and asserts on the health report it publishes plus the writes it makes. No network, no tokens.

9 assertions. 7 fail against the unfixed script, including the exact #995 shape:

FAIL  a dead Building issue is flagged as a died sidekick
FAIL  …and is NOT mis-reported as un-routed
FAIL  the issue itself gets marked dor-stuck

The two that pass pre-fix are the ones that should, which is what shows the fix is narrow: a live build stays quiet, and a genuinely un-routed issue is still flagged.

Wired into the existing ci-scripts job in pr.yml, alongside the guards added in #994.

Note on how this was routed

The DoR bug agent independently traced #995 against the current scripts, confirmed it "correct in every particular", and routed it state:out-of-pipeline — correct, since CI/automation defects go through the normal dev flow rather than the autonomous builder. Its fix sketch and regression-risk notes shaped this change; the build_phase() approach is the first of the two options it laid out.

🤖 Generated with Claude Code