daily-fix: settle whether consistency-checker runs for paren

Overview / Motivation

Auto-filed by the /daily 2026-07-26 problem sweep (route 2). Surfaced by 1 independent miner group(s) over the 2026-07-26 session transcripts.

Goal

State explicitly whether the consistency-checker runs for a parentless kind: infra task, and — if it is skipped — that the skip is recorded with that reason.

Workflow gap

• Bug observed: #1697 (kind: infra, no parent) spawned the consistency-checker, which returned PASS while recording "Parent experiment(s): N/A"; #1711 (kind: infra, no parent, same day) skipped it with the note "Consistency-checker also skipped (kind:infra, no parent experiment)" — the same task shape given two different treatments in the plan-review gate.
• Why it is a workflow gap: no rule states which treatment is correct. The one "no parent" clause that exists is experiment-shaped and gives an answer that cannot apply to an infra task, so each session decides for itself and the coverage of the gate is unpredictable.
• Confidence (emitter): high
• verified-at-filing: task state read at compose time via uv run python scripts/task.py view <N> --json | jq — #1697: kind: infra, no parent, epm:consistency markers 1; #1711: kind: infra, no parent, epm:consistency markers 0. Absence greps, per target: grep -niE 'kind: ?infra|no parent|parentless|skip' .claude/agents/consistency-checker.md → 1 hit (L334), and its context is experiment-shaped, not infra (quoted below); grep -n 'consistency-checker' .claude/skills/adversarial-planner/SKILL.md → 9 hits (L465, L472, L502, L668, L672, L804, L807, L921, L949) — none conditions the spawn on kind or on the presence of a parent; sed -n '1704,1745p' .claude/skills/issue/SKILL.md | grep -niE 'kind|parent' (Step 2b) → 3 hits, every one naming the parent recipe/task as an INPUT or as a check row ("needs only the drafted plan + the parent recipe"; "Related tasks (cited in the plan's prior work, parent task, …)"; "| Single variable change from parent | BLOCK: …") — none a precondition on kind or on the presence of a parent. Landed-fix check: git log --oneline --since='7 days ago' -- .claude/agents/consistency-checker.md → 0 commits. (2026-07-26)

Context binding. The single existing no-parent clause, .claude/agents/consistency-checker.md:334, reads: "If the experiment has no parent (first in a new direction), check against the project's standard baseline (Qwen-2.5-7B, standard eval suite)." That instruction is unexecutable for a workflow-surface infra task — there is no model, no eval suite and no recipe to compare — which is why #1697's checker fell back to recording "N/A" for every row instead. The clause exists; it does not cover this case, so the change extends it rather than duplicating it.

Evidence

• #1697, session 7df6ce4c, 2026-07-26T09:53:06Z: the checker ran and its epm:consistency note reads "## Consistency Check: #1697 vs related experiments\n\n**Verdict: PASS**\n\n### Parent experiment(s): N/A — \kind: infra` workflow-fix task, no experimental parent. Related: #1682, #1675 …"`.
• #1711, session 67cf175e, 2026-07-26T14:39:13Z: the checker was skipped — "[codex-quota-outage] codex composers skipped — quota sentinel live until 2026-08-06T13:26:00Z (#1204 pre-spawn check); single-Claude per no-show fallback for all 3 Phase-2 lens critics. Consistency-checker also skipped (kind:infra, no parent experiment)." No epm:consistency marker exists on the task (0, confirmed above).
• Measured cost: none directly. The cost is a coverage inconsistency in a plan-review gate, and — on the skip path — a missing epm:consistency marker, which makes "was this gate run?" unanswerable from the task record alone.

Proposed change

• .claude/agents/consistency-checker.md — extend the no-parent clause (L334) with the parentless non-experiment case: state whether a kind: infra task with no experimental parent is in scope, and what the checker compares against when it is (the plausible answer, for the plan to confirm or reject: the sibling/prior tasks that touched the same target file, which is precisely what #1697's checker did on its own initiative when it listed #1682, #1675, #1646, #1634, #865).
• .claude/skills/adversarial-planner/SKILL.md Phase 2 spawn list (the consistency-checker row at L921 and the spawn snippet at L804-807) — record the same precondition so the spawn decision is not re-derived per session.
• .claude/skills/issue/SKILL.md Step 2b — mirror the precondition in one line where the step describes the checker's inputs.
• Whichever way the decision goes, require the SKIP to be recorded: an epm:consistency marker (or an epm:progress note) naming the reason, so a reader of the task record can tell a deliberate skip from a gate that silently did not run. #1711's note already has the right shape — the change makes it obligatory rather than incidental.
• Prefer ONE decision over a per-session judgement call; the value of this fix is uniformity, and either answer is defensible as long as it is written down.

Scope / surfaces

• Primary target: .claude/agents/consistency-checker.md
• .claude/skills/adversarial-planner/SKILL.md (Phase 2 spawn list + spawn snippet)
• .claude/skills/issue/SKILL.md (Step 2b)

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.
• 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: fcd200fd6757

• workflow_fix_target: .claude/agents/consistency-checker.md

• fingerprint: PENDING

/daily 2026-07-26 route-2 filing. Miner refs: G-P9.