chore: the foundation says where each part of the preset contract actually lives

[YaelAnaya]

The shadcn/Nova foundation is in place and working. What was left over is
representational: three files described the preset contract in ways that were no
longer true, and one machine-readable claim had no gate in front of it. This
reconciles the description with the implementation and changes no runtime value.

The radius question, settled

DESIGN.md records the decoded preset as including radius: medium, the
stylesheet materializes it as --radius: 0.625rem, and components.json has no
radius field. The obvious repair — write the field in — is wrong, and verifying
it inside the repository's Linux container is what this change rests on.

The preset decodes to ten values, none of them a base:

$ shadcn preset decode b3bXyyPdWj --json
{
  "code": "b3bXyyPdWj",
  "version": "b",
  "values": {
    "menuColor": "inverted",
    "menuAccent": "subtle",
    "radius": "medium",
    "font": "inter",
    "iconLibrary": "tabler",
    "theme": "neutral",
    "baseColor": "neutral",
    "style": "nova",
    "chartColor": "orange",
    "fontHeading": "geist"
  },
  "derived": [],
  "url": "https://ui.shadcn.com/create?preset=b3bXyyPdWj"
}

The CLI is 4.18.0, the version DESIGN.md already names as its reference, and
the version pnpm-lock.yaml pins. Its config schema is strict, so the field is
rejected rather than absorbed — parsing the current configuration with radius
added answers:

success: false
errors: [
  {
    "code": "unrecognized_keys",
    "keys": [
      "radius"
    ],
    "path": [],
    "message": "Unrecognized key(s) in object: 'radius'"
  }
]

Adding it would not be harmless documentation; it would break every shadcn
invocation that reads the file. The configuration as committed parses cleanly and
is unchanged here. So the contract is one intent held in three layers, which is
what DESIGN.md now says:

Layer	Owns	Radius
Preset intent	What the code decodes to	radius: medium
Runtime	The value that paints, in styles.css	--radius: 0.625rem
CLI configuration	The schema-supported fields, in components.json	no field — the schema defines none

base: radix is worth one note for the next reader: it is not a decoded value at
all, but an initialization choice folded into style: "radix-nova".

The icon set, described as it is

Four places still called the primitives "Radix + lucide" — the ui-core package
description, the stack table in AGENTS.md, the frontend setup skill, and the
architecture note. The setup skill's version was the costly one, because it
instructed contributors to reach for the library the foundation had moved off.

DESIGN.md's own anti-pattern was stale in a way that misdirected: it placed the
remaining debt at screen level. No screen imports lucide-react any more. All
eleven remaining imports sit under frontend/ui-core/src/annotator/, and that is
what the rule now names.

The gate

Nothing read components.json, so the repair this change argues against had
nothing standing in its way. The new invariant in
tests/scripts/design_tokens.test.mjs asserts the key set rather than a count,
so a failure names what moved, and pins the preset values the schema does carry.
It was verified by breaking it: adding radius fails it, drifting iconLibrary
off tabler fails it, and restoring the file returns it to green.

Found, not fixed

• @visionset/app declares lucide-react and has no import of it — three
  mentions in src/demo/ToolStrip.tsx comments are the only occurrences. The
  declaration is removable, but the frontend dependency bump (chore(deps): bump the frontend-minor-patch group across 1 directory with 5 updates #768) edits that
  exact line and the lockfile, so removing it here would conflict for no gain.
  The annotator migration can drop both together.
• src/demo/ToolStrip.tsx explains its hand-drawn glyphs by saying DESIGN.md
  pins lucide-react. That has not been true since the preset landed. Left
  alone as product code outside this change's scope.

Test plan

Everything ran inside Docker/Linux; no repository validation ran natively on the
host. Two things are worth recording for whoever runs these next.

The app compose service is not a test runner: its image bakes manifests and
tsconfigs but neither eslint.config.js nor vitest.config.ts, so the suites
there fail in the thousands and lint dies on a flat-config error — the same shape
as exec api pytest exiting 5 because tests/ is not mounted. And .gitattributes
forces LF only for *.sh, so on a core.autocrlf=true checkout every Markdown
file is CRLF while the blobs are LF; copying that working tree into a container
fails roughly eighty anchor assertions in files nobody touched, because the
trailing \r corrupts every heading slug. The run below normalizes to LF, which
is what a Linux checkout gives.

A full tree with a real install, in node:24-bookworm:

install (pnpm install --frozen-lockfile)  PASS
build (pnpm -r build)                     PASS
@visionset/ui-core lint (incl. typecheck) PASS
@visionset/ui-core test                   PASS
@visionset/app build                      PASS
@visionset/app lint                       PASS
@visionset/annotator test                 PASS
pnpm test:scripts                         PASS
pnpm version:check                        PASS

The design and documentation gates were re-run against the final content after a
prose rewrap: design_tokens, docs_links and docs_sidebar together report 19
passing, 0 failing.

Playwright was not run. No runtime CSS, token, primitive or product-UI file
changes here, so there is no rendered surface for it to inspect; components.json,
styles.css and tokens.ts are byte-for-byte unchanged.