[daniele21]

MEASURE — compact insight-first model cards

Following the user-reported mobile screenshot, the cards now explain why each model matters instead of leaving quality/latency numbers floating in oversized empty space.

• An evidence-derived context label TOP QUALITY, TOP LOCAL, FASTEST LOCAL P50, or QUALITY RANK NN (within current selection). All labels are calculated at runtime from published frontier.points; no hardcoded model claims.
• Clear model name + one-sentence explanation, followed by a consistent metric strip: Quality / 100 and Median p50, each with its measured value and caption.
• Three-row compact grid, restrained vertical gaps, brand Plus Jakarta Sans + Inter, light/dark semantic tokens, selected state and mobile swipe affordance. The horizontal snap rail, filters, Show All, plot point inspection, source links and evidence limits remain intact.
• More precise Playwright measurements: first card height ≤250px, empty vertical gaps ≤20px, ≥14px role and two correctly labeled metrics. Also checks source-provenance and no page-wide overflow at 320/390/768/1440px in light/dark.

Tests

• ✅ Technical validation succeeded: Astro build, Node contract tests, smoke tests.
• ✅ Homepage visual review succeeded: includes compact model-card DOM geometry and responsive layout tests across four viewports and both themes.
• ⏳ Global visual workflow continues separately through featured project-page captures, which were outside the scope of this PR.

Per user request, merge into dev once the MEASURE-specific checks pass.