refine: propose adjustments to 1 rule(s)

[claude]

⚠️ Human review required. This PR mutates NLPM's own rulebook.
Evaluate each proposed change against the evidence. Merge the
sound ones; close the rest. Do not merge without reading
every rule section.

R15 — noisy (evidence tagged R19 in the auditor's proposal)

Current metrics: hits=4, merged/contributed=0/0, self-FP=3, maintainer-rejected=0, downstream-suppressions=0

Evidence:

• Rule-gap signals (up to 3):
  • The no-empty-input-handling rule should not penalise commands whose body is a single unconditional message.
  • The no-empty-input-handling rule should not penalise commands whose body is a single unconditional message.
  • The no-empty-input-handling rule should not penalise commands whose body is a single unconditional message.

Proposed change: The input labeled this proposal R19, but the gap text names "the no-empty-input-handling rule," which is R15 ("Handle empty input") per skills/nlpm/scoring/SKILL.md's rule-name table — R19 governs user-invocable: false, an unrelated frontmatter requirement. R19 already carries an equivalent exception ("Does not apply to: standalone commands, including commands whose body is a single unconditional message") from a prior refinement round, so no edit was needed there. Given the self-FP rate (3 of 4 hits) and the gap text, I applied the analogous exception to R15 instead: commands whose body is a single unconditional message with no branching on $ARGUMENTS have nothing to guard against, so they're now exempt from the empty-input-handling requirement. Flagging the rule_id mismatch explicitly for reviewer judgment — reject this edit if the retagging reasoning doesn't hold.

Proposals without evidence

• R31 (disputed) — no maintainer quotes, no rule_gaps. downstream_reasons has 5 entries but every reason field is literally "(no reason given)" (all threshold_adjustment type, which doesn't correspond to R31's fail-open-hooks content) — no actionable signal.
• R06 (noisy) — empty quotes, empty rule_gaps, empty downstream_reasons. Metrics alone (verify_rate=0.077, introduced=7) don't indicate which wording to tighten.
• R15 (noisy, filed as its own proposal separate from the R19-tagged one above) — empty quotes, empty rule_gaps, empty downstream_reasons.
• R16 (noisy) — empty quotes, empty rule_gaps, empty downstream_reasons.

Not applicable to this file

• SEC-unpinned-semver (disputed) — has quotes and a rule_gap, but this ID isn't part of the R01–R50 catalog in skills/nlpm/rules/SKILL.md; it's a security-scanner pattern (skills/nlpm/security/SKILL.md). Per site/how-it-evolves.md, this was already "demoted" via rule-health.py classifier logic in v0.7.22, not via rulebook wording — out of scope for this PR, which may only touch skills/nlpm/rules/SKILL.md.
• CC-stale-count (disputed) — same reason: not an R-numbered rule (it's a cross-artifact-consistency checker pattern), and was already cleared via the applied_separately reclassification fix in rule-health.py (v0.7.22).
• SEC-hardcoded-api-key (noisy) — has a strong, actionable rule_gap (exempt public client-side analytics keys — PostHog/Segment/Mixpanel — from hardcoded-credential findings), but this pattern lives in skills/nlpm/security/SKILL.md, not the rules catalog. Recommend routing this fix to a security-skill-focused refinement pass; no edit made here.

[github-actions]

🔍 Preview deploy

Preview URL: https://nlpm.com/_pr/817/

Built from 18c3703 at 2026-08-09T09:22:40.563Z.

It may take 30–60s for the GitHub Pages CDN to serve the new content. The preview is removed automatically when the PR closes.