feat(viewer): approaches view — offline synthesis of strategies tried

[aktasbatuhan]

Viewer roadmap item D. Adds an approaches view at /setup/{label}/run/{idx}/approaches, modeled on AlphaEvolve's AI-overview tab — but built as a pure offline disk reader (no LLM at render time, consistent with the rest of the viewer).

Two sections

1. "The ideas that worked" — the new-best lineage. Each step that set a new best, with its score, Δ, and the first sentence of its hypothesis. Read top-to-bottom, it's the chain of ideas that actually moved the score (e.g. for autocorrelation: L-BFGS-B → decreasing-sequence init → differential evolution → smooth-max approximation, 0.13 → 0.45).
2. "Techniques the AI kept trying" — salient technique phrases repeated across the HMRD notes, ranked by how often a phrase coincided with a new best, then by breadth. Prefers 2-word phrases (named techniques like "local search", "multi-start slsqp"), drops generic standalone words and unigrams already subsumed by a kept phrase.

Honest by construction

Framed in-page as a term map, not semantic clustering. Degrades gracefully with explicit empty-states when a run left no notes, no step improved on the seed, or phrasing didn't repeat — e.g. circle-packing runs whose initial program is already near-optimal correctly show "No step improved on the starting program's score" rather than a fabricated lineage.

Verification

• Helpers (_first_sentence, _extract_techniques) import without the viewer extra and are unit-tested directly (ranking, generic-word/singleton filtering, sentence truncation); an app test covers the page + dashboard link.
• Full suite green (322 tests). Live crawl: approaches route 200 across every arm in both tasks; verified visually on auto (rich lineage) and circle (graceful empty-state).

🤖 Generated with Claude Code