Release v2.1.0

[vicmaster]

Version bump only — package.json, the lockfile, and the McpServer version in src/index.ts. All three verified to agree; npm run build and npm run test:fast (68/68) pass on the bump.

What ships

Phase 29 — commerce fluency. Closed with a literal acceptance proof: the checkout screen that started the phase was rebuilt from the same inputs with none of its four workarounds and scored 100/100 with a clean stress run in both themes, against a 77 baseline.

• A commerce genre, so a transactional screen is no longer penalised for showing honest prices — the tell that cost 20 points on a checkout page.
• Field-wise typography preservation, so a partially-specified inherited token no longer silently strips the generated personality's font family, and a designSystemConflicts report for inherited tokens the generator's own vocabulary owns.
• Every component scaffold now references type roles rather than literal pixel sizes, and survives long-text and i18n content.
• Advisory honesty: the unsatisfiable sibling-padding check removed, spacing variety measured against the canvas's own declared scale, and readiness no longer withheld when nothing blocking remains.
• screenshot gained fullPage for designs taller than their artboard, and responsive: "fixed" was withdrawn.

Continuous integration. The project had 102 test files and no way to run them as a suite. There is now npm test, and every pull request runs build plus tests on Node 22 and 24 alongside a browser job. It found three tests already broken, two of them red at the v2.0.0 tag, and then caught two real scaffold-overflow bugs that only reproduce on Linux font metrics.

Viewer performance. The worst project page went from 4.2 seconds to 0.09 on a real 178-canvas store. The gallery was recomputing a score cache key per card, and that key re-listed and re-hashed every canvas in the store — quadratic in the whole store rather than the project being viewed.

Why minor and not major

The only withdrawal in this range is responsive: "fixed", and it was a documented no-op: the renderer never read it. I checked rather than assumed — passing it today still succeeds, still renders, and canvases carrying it still load unchanged. The package declares no types or exports field, so there is no typed library surface for a consumer to compile against. Nothing user-visible breaks.

Scores do move on some canvases, because the sibling-padding advisory that could not be satisfied is gone. They move up, and the benchmark was re-baselined in #203.