Functional client-side sandbox comparison site

[chrisns]

A static, client-side GitHub Pages page that presents the probe's reports so someone can (a) compare AI-agent sandboxes to make a trust decision, (b) track degradations/improvements as harness and supporting-tech versions move, and (c) absorb new harnesses with no code change. Functional pass only — styling is deliberately plain, ready to hand to design.

What's here

• site/ — index.html + vanilla app.js + placeholder style.css. All derivation (fingerprint-collapse, baseline-normalization, categorization, flips, exposure) is client-side from one concatenated reports file; falls back to sample-data.json in dev.
• Views: baseline-normalized three-state matrix (🟥 leaked / 🟩 blocked / ⬜ n/a) with enforcement + root badges and ▲/▼ change markers; exposure-over-time ECharts step-line (CDN + SRI, progressive enhancement) with version-release markers and multi-identity overlay; per-capability heatmap; attributed flip-log; cell drill-down.
• scripts/gen-sample-data.mjs — generates the dev fixture (history with real regressions/improvements to exercise every view).
• Docs: CONTEXT.md (glossary), docs/adr/0001-* (pipeline decision + why not Actions artifacts), docs/reporting-site-plan.md (build checklist).

Extendability (holds by construction)

• New harness → new row/line, automatic (identity from data).
• New finding type → mapped category or an Other column, never dropped.

Not in this PR (tracked in the plan doc)

• Track 1–3: probe list-targets, the soft-seeder, and the scan-matrix aggregate/publish job. The site is developed against the fixture until those land.
• The design system.

https://claude.ai/code/session_01PtFjYK8MdnDMmFYVWPDRrn

[chrisns]

Validated against real probe data, not just the fixture: seeded a throwaway $HOME, ran a real `direct` baseline and real Seatbelt-confined scans (`sandbox-exec`), and pointed the page at the resulting `all-reports.json`. Seatbelt cut sensitive reads 31→5 and (in the strict config) zeroed network — the page rendered the real baseline-vs-sandbox diff correctly.

That exercise caught a renderer bug now fixed here (bf1451e): a task that runs and finds nothing emits the finding type with an empty value (e.g. external_host_dns_resolution: []), which was being counted as a leak. Gated on a non-empty value — blocked capabilities now render as blocked. The data pipeline that produces real all-reports.json is #41.