[mvilacad]

Closes #571

Part 1 of the Concepts program: the contract that keeps claims about users honest.

What ships

• Methodology page /docs/concepts/evidence/: five claim-level strength labels (established, supported, preliminary, practice, heuristic), nine rules, how to read the atlas, how to validate, a Limits box with a review date. No "science" branding: the section is Evidence.
• References dataset site/src/data/evidence/references.json (45 sources). Every one was opened on verified.date, with the wording relied on kept as verified.excerpt and the route used in verified.how. All 45 are claim-read.
• <Claim>, <Limits>, <Refs> docs components and the concepts index with the nine-part page template.
• Markdown twins keep strength, scope, limits and "does not support" (/llms/concepts/*.txt).
• Tests (packages/ui/tests/evidence-rules.test.ts): banned phrases, a kind caps the strength (a standard is practice, a vendor study preliminary), every cited source exists and was read, every claim has scope and sources, every concept has a dated Limits box and References, checklists are plain lists.

Fail before, pass after

• has read the wording of every source it records first failed on E07 (excerpt was only the page title). Re-read the WCAG quick reference, recorded the 1.4.10, 2.4.11 and 2.5.8 wording: passes.
• E02 and E18 were record-only. Read their abstracts (author-hosted PDF; arXiv version of the survey, stated in how): both claim-read, so the old "identity only" escape hatch and its vacuous test are gone.
• E01 excerpt now holds the 17 ms wording the concept pages cite.

Verified

bun test packages/ui/tests/evidence-rules.test.ts: pass. Pre-push validate: ok.

Left open

Nothing for this layer. #578 (atlas) and #586 (concepts) are stacked on this branch.