torque-loop: Skill for Claude Code

/preflight — embarrass the branch before the reviewer does

v0.7 shipped, got reviewed, and needed a same-day PATCH-THEN-KEEP hardening pass (fog write bypassed on --json, probe invariants convention-only, gated reasons unenforced). Every one was findable before the PR. This skill is that finding pass.

The mechanizable checks live in a script, not in prose — a discipline a model has to remember to run is weak. scripts/preflight.js owns the deterministic verdicts; this skill runs it, then does the judgment the script deliberately leaves open.

Step 1 — Run the mechanical pass

npm run preflight          # or: node scripts/preflight.js [base-ref]   (default base: main)

It prints all 12 checks in a fixed shape and exits non-zero if any MECHANICAL check fails. Trust its verdict on these five — they need no judgment:

• 1 green world — npm test + ratchet doctor both exit 0.
• 4 version alignment — five version fields + README examples all match.
• 7 dependency gate — no new dependencies/devDependencies (or a Danny quote in [Unreleased]).
• 9 leak scan — no private paths (reference/PROBLEM-STATEMENT-*, *.private.md, .lucid/, .ratchet/, .sandbox-*) and no exact private-line quotes in the diff.
• 10 trace tags — changed durable docs (CLAUDE.md, reference/*, handoffs) carry a Traced by: tag; commits carry a Co-authored-by footer. templates/ is exempt.

A MECHANICAL FAIL is real; fix it before anything else. The script's SMALLEST PATCHES block names each fix.

Step 2 — Rule on the checks the script only gathers evidence for

The script prints these as HUMAN: lines with the candidates it found. You make the call — it cannot. Read the gathered evidence, then decide PASS/FAIL for each:

• 2 testless change (weak signal — distrust a clean line). The script's token-grep is noisy: common identifiers match half the suite, so "has test hits" is near-meaningless. For every src/|bin/ hunk, name the specific test that fails without it. No name → FAIL.
• 3 weakened falsifier. The script flags removed asserts / newly-loose lines in test/. Decide whether any assertion was genuinely loosened without a commit-body reason. If so → FAIL.
• 5 prose enforcement. The script lists new must/never/requires lines in SKILL.md/README. For each, point at the CLI line that enforces it + the test proving the refusal throws, OR confirm the CHANGELOG labels it prompt-level. Unlabeled invariant → FAIL.
• 6 diff minimality. The script catches renames + whitespace-only files (the cheap cases) and is blind to elegance-churn (the expensive one). Map every remaining hunk to the locked target yourself. Any hunk that doesn't map → FAIL: revert this hunk.
• 8 parked-decision creep. The script lists parked loop ids and any the diff mentions. Decide whether a hunk actually implements a parked loop (a mention alone is fine). If so → FAIL.
• 11 scope + emptiness (weak signal). For any new score/read the script surfaced, confirm it exposes a scope: field and that new rendered sections state emptiness.
• 12 changelog. The script shows whether [Unreleased] was touched and whether a CLI-enforced/prompt-level label is present. Decide if behavior changed and the entry is adequate.