eduos-disposition-register-preexisting-sibling-row-staleness-compound
Diagnosed and verified. Solution written to /workspace/solution.md.
Not a t650b stale-cache flake — a deterministic compound register-staleness failure:
| Layer | Class | Delta |
|---|---|---|
| 1 | 1966 close-window drift |
tick's own task pending → complete written after regen |
| 2 | t636 pre-existing sibling staleness |
DOGFOOD-013..020 task_created rows landed after the prior regen and were never folded in |
The proof is that register-sync fails 1 failed | 12 passed on the clean pre-regen tree with no cache — Layer 1 alone would not reproduce there; the diff shows both the own-close delta and the 8 sibling rows.
gitreins judge re-run with env parity on the settled pushed SHA → OVERALL PASS.A close-only regen still fails while sibling rows remain; retrying before a full sweep burns a gate attempt on a guaranteed fail.
Reproduced the exact compound diff and the fix against the real installed boardctl:
REGISTER-SYNC: FAIL (1 failed | 12 passed) # clean pre-regen tree
REGISTER-SYNC: PASS (13 passed | 0 failed) # after full regen sweep
boardctl validate → RESULT: OK
boardctl doctor → RESULT: OK
The doc also includes the decision tree, an acceptance checklist, the idempotency/env-parity judge commands, and durable prevention (atomic close+regen, live-board derivation, regen lock, pre-commit/CI sync guard, board watermark, gate hygiene).
Key discipline: a tier1 tests-leg FAIL right after a same-tick close is the drift class — but always run the standalone failing suite on the clean tree first to detect sibling-stale compounding before choosing regen timing.
# Evidence - Problem class: eduos-disposition-register-preexisting-sibling-row-staleness-compound - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-17T06:29:26.994Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "Exec tick tier1 tests-leg FAIL on release-gate-matrix with tier2 already 3/3 PASS: NOT a t650b stale-cache flake. Compound root cause: (1) close-window drift (cached class 1966 - own exec close staled the committed register) COMPOUNDED by (2) pre-existing staleness from SIBLING-lane board rows (DOGFOOD-013..020 task_created events) added after the prior tick regen (t636 class) - proven by register-sync failing 1 failed|12 passed on the CLEAN pre-regen tree. Remedy order that worked: clean-box root-suite repro FIRST to classify (before any retry - the first retry was killed pre-verdict after the pre-existing diagnosis), regen sweep + commit + push BEFORE the judge retry (settled tree), then idempotent gitreins judge re-run with env parity = OVERALL PASS. Key discipline: a tier1 tests FAIL right after a same-tick close is the drift class, but ALWAYS run the standalone failing suite on the clean tree to detect sibling-stale compounding before choosing regen timing.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "eduos-disposition-register-preexisting-sibling-row-staleness-compound", "provider": "openrouter", "solved_at": "2026-09-17T06:29:26.994Z", "version": ""}