feat(requirements): add PRD-calibration checkpoint for numeric-ceiling FRs (v2.18.5)

[pitimon]

Summary

Closes #237 — action item from lesson 2026-05-24-v218-4-skill-authoring-double-rescue.md §5 catching FR-007 (ADR-020) PRD-vs-reality drift before merge in v2.18.4.

Adds a new sub-step 4a. Calibrate numeric ceilings against precedent to skills/requirements/SKILL.md Process step 4. When an EARS criterion sets an upper bound on lines/words/characters of a markdown artifact (ADR, guide, skill), the skill now nudges the author to:

1. Identify the closest precedent in repo by artifact type (docs/adr/ADR-0*.md / guides/*.md / skills/*/SKILL.md)
2. Measure body excluding YAML frontmatter: awk '/^---$/{c++; next} c>=2' <file> | wc -l
3. Set the FR ceiling at precedent_max × 1.20, because aspirational round numbers contaminate /consistency-check runs once the artifact lands at its actual required size
4. Opt out with a one-line FR-comment rationale when no precedent exists or the cap is set by a different constraint (hook token budget, validator string limit, etc.)

Case study: v2.18.4 FR-007 in docs/specs/skill-authoring-guide-235/prd.md shipped at ≤50 lines body against an ADR-017 precedent of ~150 lines and required pre-merge amendment to ≤150. @8-habit-reviewer caught the drift reactively; this checkpoint moves the catch upstream into the authoring step itself (H1 + H5 + H7).

Acceptance criteria check (per issue #237)

• AC1 — skills/requirements/SKILL.md step 4a includes the 4-step sequence (skills/requirements/SKILL.md:78-82)
• AC2 — FR-007 case study cited with one-line precedent example (skills/requirements/SKILL.md:81)
• AC3 — Opt-out clause covers novel-artifact + different-constraint rationales (skills/requirements/SKILL.md:82)
• AC4 — 7 lines added (well within ≤15 budget); Check 25 description ≤1024 chars + F1 word budget preserved
• AC5 — Check 26 NEVER/MUST hygiene satisfied via because reason marker (skills/requirements/SKILL.md:81)
• AC6 — guides/templates/prd-template.md Success Criteria section shows calibrated-vs-uncalibrated FR-007 example

Out of scope (per issue #237)

• No validator extension, no PreToolUse hook — runtime enforcement belongs in pitimon/claude-governance per plugin boundary (memory obs #233270)
• No retroactive PRD audit — existing PRDs in docs/specs/ not touched
• No extension to non-numeric FR fields — the friction case is specifically numeric ceilings on markdown artifacts

Pre-PR cross-verify (@8-habit-reviewer)

16/17 PASS + 1 LOW (meta-honesty) + 1 N/A (Q13 solo single-skill edit).

The single LOW finding is the kind of recursive honesty this checkpoint preaches: issue #237's own AC #4 (≤15-line cap) was itself an uncalibrated round number. The closest precedent for "sub-step addition to a SKILL.md Process section" is the EARS opt-out paragraph added in v2.5.0 (~4 lines); a strict +20% ceiling would be ~5 lines. The actual implementation landed at 7 lines, slightly over a v2.5.0-derived ceiling.

This does not block merge, because the contamination mechanism only fires when ACs are persisted as spec FRs that /consistency-check reads — issue body ACs are not persisted artifacts. The H8 modeling discipline is to acknowledge the gap rather than gloss it over: AC #4's ≤15-line cap was aspirational; future issue-AC-budget proposals should themselves apply step 4a or declare the opt-out rationale.

Consumer-doctrine bump (per ADR-019)

skills/** + guides/** edits MUST bump per Check 27. Patch grain v2.18.4 → v2.18.5 atomic across:

• .claude-plugin/plugin.json + .claude-plugin/marketplace.json
• README.md (badge + footer + What's New section)
• SELF-CHECK.md (header + per-release row + footer)

Forward-Guardrail Sunset 2026-11-24

Per ADR-017 mechanism (not ADR-016 eviction). Reversal criteria: no /reflect cite, no calibrated FR in subsequent persisted PRDs, FR-007 case remains the only friction signal. If sunset triggers, revert sub-step 4a; the case study + template example remain as authoring lore.

Test plan

• tests/validate-structure.sh — 358 PASS / 0 FAIL (Check 25, 26, 27 all pass)
• tests/validate-content.sh — 275 PASS / 0 FAIL / 1 WARN
• tests/test-skill-graph.sh — 75 PASS / 0 FAIL
• tests/test-verbosity-hook.sh — 19 PASS / 0 FAIL
• CI green on PR (link-check + validator suite)
• Post-merge: tag v2.18.5 + GitHub release with this body as release notes

Refs #237