v1.2.0 — Public benchmark explorer

Add anonymous benchmark browsing, canonical model profiles, source-label discovery, full-server-catalog typeahead, and URL-persisted comparison of 2–4 models or exact unresolved source rows. Public reads bypass JWT/project probes; existing authentication, CI setup and project flows remain compatible.

Evidence keeps source/version/protocol/evaluator/snapshot/metric/coverage scopes separate. Users choose among multiple observations; no automatic maximum, composite score, inferred identity or runtime-support badge. Missing values remain missing. Shared benchmark explanations support focus, Escape, hover and touch. The explorer consumes the additive backend 1.4.0 public metadata contract.

Validation: 132 unit/component tests, 21 HTTP integration tests, 12 new desktop/mobile browser checks, 12 existing public/auth browser checks and 6 Google-auth browser checks. Typecheck, lint (no errors), production build and whitespace checks pass. Local real-API smoke covered all 19 frozen families, 13 CodeReviewBench rows, unresolved comparison/reload and narrow screens. Self-review fixes preserve comparison URLs, correct popover focus, keep home single-screen, and wrap long frozen revisions.

Release: 1.1.0 → 1.2.0 (MINOR) for backward-compatible public explorer functionality. Publish to GHCR only; deployment and production imports remain separate.