feat(ui): compare runs across workspaces on two axes, by case, and score by score - #42

Summary

Comparing the same experiment across several workspaces (for example 7 ASR rephrase products) had no entry point, labelled every column the same, merged the matrix rows of a case, and drew only a bar chart without uncertainty. This PR adds a cross-workspace comparison and charts that tell a real gap from noise.

vieval

• defineScore(name, { rank }) replaces primary: rank 1 is the headline, 2 the next core score, and so on. score.defined adds rank and keeps primary (rank 1), so earlier reports still open. Docs updated.

Server (/api/compare/*)

• Rows split by the matrix axes every column has (variant, such as context=none); an axis only one column runs (a prompt axis in one workspace) is averaged into a pooled cell, so columns stay paired case by case.
• Cells carry counted scores (a failed case counts 0), numeric metrics, and durationMs; columns carry workspaceId and project; the response carries metricDefinitions.

Viewer

• Home and Comparisons list: experiments run in several workspaces as leaderboards; the list keeps only comparisons of two or more things.
• Compare page:
  • Y / X / baseline pickers, each axis previewed beside the menu (a drawer on phones);
  • Two-axis distribution: means with 95% Student-t intervals, Pareto front, crosshair readout, linear/log per axis, hover preview and click-pinned card (use as baseline, hide, open run);
  • Differences from the baseline paired by case (estimation plot) with hover values for the mean and interval ends;
  • Every score and metric as compact heat cells shared with the matrix (HeatCell), best per axis outlined, a beyond-noise tally per column;
  • columns can be hidden from the charts (the plane zooms, the table fades the row);
  • two or three chart columns on wide screens.
• Scores are ordered by rank everywhere; losses use the theme hue turned 180° (--series-1-opposite); heat fills stay within 50% (measured ≥ 5.7:1 contrast), also in the compact matrix.

Testing

• pnpm lint, pnpm typecheck, pnpm exec vitest run: 86 files, 712 tests passed.
• New tests: comparison rows split by shared axes and pooled (regression), cell metrics, rank (DSL and event), orderScores, comparison.ts (intervals, paired deltas, Pareto front, labels).
• Checked in the browser against real reports (7 workspaces) at 390, 1000, 1440, 2560 and 3840 px, dark and light.

Follow-ups

• Extract a shared virtual table from TaskPage / JsonlTable and use it for the compare cases table.
• A 90 / 95 / 99% confidence option on the charts.
• Within-noise tie-break by the next rank.

🤖 Generated with Claude Code