daily-fix: closed-sibling probe suppressed 21 of 24 filings

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by the sweep's own filing run — the probe blocked 21 of this run's 24 filings.

The #1711 closed-sibling pre-filing probe and the #1674 landed-fix probe, both of which merged on 2026-07-26, together record a terminal non-filing outcome for the large majority of a nightly batch on this repo's hot workflow files. Tonight the operator eyeballed all 21 and every one was a false positive, so the batch was filed with --retry-suspects. A run that did not check would have silently dropped 21 verified defects.

Goal

Tighten the closed-sibling / landed-fix suspect predicates so a shared hot target file or a generic title word does not by itself record a non-filing outcome, and make the suppressed count impossible to miss.

Workflow gap

• Bug observed: on the 2026-07-26 batch, daily_drive_filings.py --dry-run recorded CLOSED-SIBLING-SUSPECT or LANDED-FIX-SUSPECT (terminal, NOT filing) for 21 of 24 manifest items; only 3 would have filed. All 21 were manually confirmed unrelated to the cited sibling.
• Why it is a workflow gap: the nightly sweep concentrates on a handful of hot workflow-surface files (.claude/skills/issue/SKILL.md, scripts/workflow_lint.py, .claude/agents/*.md), so within any 7-day window nearly every new filing shares a target file with some recently-closed sibling. A matched: target-only hit therefore carries almost no evidence of duplication on this corpus, yet it suppresses the filing.
• Confidence (emitter): high
• verified-at-filing: uv run python scripts/daily_drive_filings.py --dir logs/daily/filings-2026-07-26 --dry-run → 21 suspect lines vs 3 FILE lines (24 items); grep -oE '\(matched: [^)]*\)' /tmp/dryrun2.log | sed 's/; .*//' | sort | uniq -c | sort -rn → top reasons title:main (9), bare target (7), title:runs (5), title:merge (4), title:tests (3), title:step (3), title:probe (3), title:path (3), title:check (3); git rev-parse --verify --quiet '9106aaf478^{commit}' → 9106aaf478... resolves, and git show 9106aaf478 --stat confirms it is task #1560, the commit that BUILT the Step 5a/10d spec-freshness sync (2026-07-20) — i.e. the mechanism tonight's spec-freshness-sync-inconsistent-tree item reports a defect IN, not a prior fix for it. (2026-07-26)

Evidence

• Representative false positives, each a terminal non-filing outcome:
  • main-red-inline-round-duty-mirror → #1648 "lint bare git-commit recipes lacking pathspec" (matched: target). #1648 is unrelated to a deleted lint function; this is the URGENT main-red restore.
  • code-reviewer-ruff-policy-pin → #1693 "phase idempotency + inter-phase schema assert" (matched: target).
  • planner-grounding-discipline → #1604 "wire identity+bias baseline + kNN retrieval mandate into planner/critic lenses" (matched: target, title:planner).
  • select-step9c-tests-ergonomics → #1697 "--map-files timeout dispersion margin" — same file and even the same flag name, but #1697 changed a timeout constant while this item is about --json being silently ignored.
  • spec-freshness-sync-inconsistent-tree → landed-fix suspect 9106aaf478, which is the commit that introduced the mechanism being reported as defective.
• Nine separate items matched on the single title token main — a word that appears in ordinary workflow prose (origin/main, "main tree", "main-red") and carries no duplication signal.
• The #1399 advisory this probe descends from was explicitly specified as advisory only, never a block (.claude/rules/workflow-fix-on-bug.md, § Recently-closed-sibling ADVISORY: "it never blocks the filing, never changes exit codes"). The #1711 probe records a terminal ledger outcome instead, which is a stronger contract than the rule documents.
• Cost tonight: zero filings lost, because the operator read the dry-run and re-ran with --retry-suspects. The exposure is the counterfactual — a run that trusts the default drops the batch silently, and the ledger reads as a clean pass.

Proposed change

• Require MORE than a bare target-file match before recording a suspect outcome: e.g. a shared target PLUS at least one shared informative title token, or a body-level overlap signal.
• Extend the non-informative token stoplist used by the title-overlap arm with the generic workflow vocabulary observed here — main, runs, step, merge, tests, path, check, probe, state, daily, zero, repo, shared — so a single such token cannot produce a match. (task_workflow.informative_title_tokens is the shared implementation; check whether the stoplist belongs there or driver-side.)
• Reconcile the contract with .claude/rules/workflow-fix-on-bug.md: either soften the probes to advisory (print, then file) to match the documented #1399 behaviour, or update the rule to state that the closed-sibling probe is now blocking-by-default with --retry-suspects as the override. Today the rule and the code disagree.
• Make the suppression legible regardless: print a one-line summary (N of M items suppressed as suspects — re-run with --retry-suspects to file) at the end of the run, and record the suppressed count in the filed.jsonl run record, so a future operator cannot mistake a mass suppression for a clean batch.

Scope / surfaces

• Primary target: scripts/daily_drive_filings.py
• src/explore_persona_space/task_workflow.py (informative_title_tokens stoplist, if the shared helper is the right home)
• .claude/rules/workflow-fix-on-bug.md (advisory-vs-blocking contract reconciliation)

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• uv run python scripts/workflow_lint.py passes (no-flags); ruff clean on touched files.
• Loosening the predicate must not reintroduce the #1350/#1329 duplicate-filing class the probe exists to prevent — a genuine same-bug sibling must still be caught. The two measured reference points: #1350 (filed 25 min after #1329 merged the same fix) should still match; tonight's 21 should not.
• This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

• fingerprint: 24883d891330

• workflow_fix_target: scripts/daily_drive_filings.py

• fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Observed directly in this run's own filing dry-run.