test: eval + bench for wideshot heuristic, ocd skill ablation study

[undeemed]

Adds test/eval infrastructure and an empirical study of whether the ocd skill actually changes outcomes. No changes to the shipped skills themselves.

What's in it

1. skills/wideshot/scripts/test/ - heuristic eval + bench

• 12-case correctness eval for pick-container.js, one case per claim/guardrail (landmark match, 3x-area rung, sibling requirement, viewport ceiling, shadow DOM, display:contents, position:fixed, scrollport clamp, page coords, bad selector, no-ancestor fallback, generic fallback).
• Runs in real headless Chrome via puppeteer-core - jsdom has no layout engine, so its rects are 0x0 and every geometric assertion would be vacuous.
• Perf bench: heuristic p50/p95 (sub-ms) + wideshot.mjs end-to-end (<2s claim) + a framing-strategy comparison (element-only vs wideshot vs full-page).

2. eval/ablation/ - controlled A/B study of the ocd skill

• Same audit task run with and without the skill, scored by deterministic render-and-measure scripts, not model judgment. Scorers are self-tested (buggy fails, reference-fix passes, symptom-patch is caught) before any agent runs.
• 5 rounds, 48+ subagent runs, code + visual + discovery + a real heavy landing page (HTML5 UP "Stellar").
• Robust finding: no correctness lift when a bug is named, visible on sight, or readable in source - even buried in a real heavy page - at 2-3x the token cost. The one real win is discovery of a source-invisible, state-dependent layout shift: base 0/8 vs ocd 8/8, with before/after screenshots.

3. README.md

• wideshot perf/comparison table.
• "Does it actually help? (measured)" section presenting the ablation evidence with honest scope.

🤖 Generated with Claude Code