[Jammy2211]

Inference evidence now opens through project → dataset family → model setup links in new tabs. Each setup page puts the qualified baseline and cost/start conditions first, with sampler, historical, parity and provenance details beneath. Direct links retain freshness, failed-refresh and local-preview qualifications; missing accepted baselines stay explicit.

Scripts Changed

• Board renderer and setup browser reuse shared dashboard sizing, hero and orchestration components; include responsive metrics, bounded images, safe deep links and no-JavaScript disclosures.
• Registry now ingests the merged producer's inference-summary@2 catalogue, with pinned receipts/snapshots for 56 records across 11 declared setups. Historical unknowns remain unknown.
• Public sampler candidates distinguish paper expectations from integration and measured evidence, with editable copyable investigation prompts and source links.
• Required Chromium CI checks navigation, new tabs, history, mobile width, clipboard fallback and no-JavaScript behavior.

Validation checklist

• 168 tests; Ruff/format; offline exchange/feed checks; real Chromium fixture and actual producer capture checks.
• Independent CLEAN review, including failed-refresh/stale/local qualification on direct pages and primary-source candidate handoff.
• Producer PR21 merged; successful publication and refresh dispatch. Captured de37acfe catalogue bytes verified independently.
• Heart STALE only: release validation incomplete: no rehearsal for current source; no RED/YELLOW reasons.

No new science measurements or accepted baselines are invented. Current evidence has no selected accepted baseline or portable prepared problem. This provides the navigation and contract for future completed baselines/experiments; it does not launch them.

Closes #15. Human-authorized full-phase completion, including merge on every green CI job and post-merge Pages verification.