feat(web): tests section, headline charts, model verdicts, methodology from data

[VincentShipsIt]

Closes #7

Summary

• Removes the .instrument-ground scanline layer that made prose pages hard to read, keeping the two radial glows; widens prose measure/size across the methodology page, empty state, and run judgment/report views.
• Rewrites /methodology/ to render content/methodology.md (hand-authored) plus generated sections built from live data: the suite table from tasks/**/task.yaml via a new loadTasks() loader, the price table with sources, the configured-vs-actual judge panel, and a "known limits" list derived from the release's own numbers (attempt count, judge families that actually judged, agents that never ran). Deletes lib/design-doc.ts.
• Adds a /tests/ section: an index of six task cards (question, category, scoring badges, current leader) and a /tests/[taskSlug]/ detail page with the verbatim prompt, fixture tree (+ inlined key files), plain-English gate sentences, hidden-test-case counts (never the spec contents), the full rubric, and every release's results for that task. Linked from the header nav, the leaderboard's category headers, and the release-scoped task gallery page.
• Home page: three ranked bar charts (score, cost, wall time) above the standings table, a one-sentence data-driven verdict (lib/verdict.ts), an upgraded score-vs-cost scatter (cheap-and-good quadrant, Pareto frontier, per-vendor colors, legend), and a changelog feed built from releases and task authored dates.
• Model page: four ranked tiles (overall, cost efficiency, wall time, telemetry), a two-sentence verdict, and a tokens-by-class stacked bar sized by dollar contribution using data/prices.json.
• Release page and its OG image now show the same verdict sentence.

Schema note

No schema changes were needed — TaskManifest, Release, and PriceTable already carried everything these pages needed.

Test plan

• bun run build (real data/) — all routes, including new /tests/ and /tests/[taskSlug]/, build and prerender.
• bun run build:sample (BENCH_DATA_DIR=./fixtures/sample-data) — confirms the methodology and tests pages still populate from the real tasks/ dir even when the release data is synthetic.
• bun run lint (biome) clean.
• Verified rendered output: verdict sentence matches the release data exactly, hidden-test spec contents never appear in the built HTML, never-run agent (Grok) renders sane ranks/tiles/token bar instead of crashing, and the release OG image PNG renders with the verdict subline.
• Not verified here (host runs tests/typecheck on Mac Studio, not this machine): bun run typecheck, bun test.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
benchmark.shipshit.dev	Ready	Preview	Sep 6, 2026 2:38pm UTC