audit: recalibrate human scoring rubric — demanding & graduated (U2)

[RasputinKaiser]

Recalibrates the plugin-audit human scoring rubric (skills/plugin-audit/references/scoring-rubric.md) so it discriminates across the quality range instead of saturating. Part of the audit-recalibration batch (Unit U2).

What changed

• Graduated, demanding criteria for all six dimensions: validity now earns the MIDDLE of each sub-point's range; full marks require excellence beyond mere validity (leaner bodies well under budget, complete/accurate publisher metadata, runtime-correct hooks, precise triggers). Credit described as a qualitative gradient — no coefficients hardcoded, so the doc stays in sync with score.py.
• "Not applicable → redistribute" rule added near the top: a genuinely-inapplicable dimension or sub-point is dropped and its weight spread proportionally, renormalizing to 100 — never a free max. Includes the no-hooks example; single-harness parity and no-risky-skills sub-points now drop-and-redistribute within their dimension.
• Frozen 5-band scale replaces the old bands: 92–100 Exceptional · 82–91 Strong · 68–81 Solid · 50–67 Needs work · <50 Poor, with one line of anti-saturation guidance (typical good-but-improvable plugin lands Solid/Strong; Exceptional reserved for near-perfect across all applicable dimensions).

Invariants held

• Six dimensions and nominal maxes unchanged: Manifest 15 · Skill 25 · Trigger 20 · Context 20 · Hooks 10 · Distribution 10 = 100. Sub-point maxes still sum to each dimension.
• Dual-harness-neutral wording preserved. No versions/CHANGELOG/README/manifests/scripts/other skills touched — only scoring-rubric.md.

Verification

• python3 scripts/validate.py → PASS 11/11, exit 0.
• Code review (agent): dimension sums = 100, redistribution rule internally consistent, frozen bands verbatim, all relative links resolve. No findings.

🤖 Generated with Claude Code