benchmark: live-mode scoring benchmark for src/scoring.js (LLM judge layer)

[devswha]

Context

tests/quality/README.md ("What it does NOT measure") notes that the deterministic quality benchmark intentionally excludes LLM-based scoring (src/scoring.js) because it is non-deterministic and adds API cost/latency, and that "a separate live-mode benchmark would be its own follow-up". This was the only follow-up note in the repo without a tracking issue — filing it so it doesn't get lost.

Scope

• An opt-in, live-mode benchmark that exercises src/scoring.js end-to-end against the suspect-zones / rebaseline corpora and reports score distributions per pattern pack.
• Complements (does not replace) the deterministic layer in tests/quality/; same opt-in posture as tests/quality/live-quality.mjs.
• Related research follow-ups: research: cross-judge agreement matrix (Claude vs Gemini vs GPT judge) #158 (cross-judge agreement matrix), research: blinded human evaluation panel — does score-low text actually read human? #159 (blinded human evaluation panel) — a live scoring benchmark is a natural substrate for both.

Non-goals

• No CI gate: live results are diagnostics, consistent with the existing live-quality stance.

Source: tests/quality/README.md follow-up note, surfaced during the 2026-06-11 remaining-work audit.