feat(cli): launch-design-implement — build a Claude Design canvas true to the prototype

[4lgn]

What

A new user-typed skill, launch-design-implement, that takes a Claude Design export to a built page in one invocation — drop the zip, name the canvas, type the command — and builds it true to the prototype under a written fidelity protocol.

The existing on-ramp (launch-design-handoff, ADR-0024) files a design and routes it onto the rail but deliberately never builds. The build half was unowned, so "take this canvas and implement it" happened ad hoc and missed the mark in six repeatable ways: values read off screenshots instead of the prototype's code; only the first render built, the states behind clicks skipped; the design normalized toward existing components and tokens; no side-by-side at the viewport the design was approved at; the wrong canvas picked from an export that holds several; data the prototype shows dropped or faked silently. Each now has a rule against it.

How it works

/launch-design-implement "<canvas>" with the export dropped in the session (or a package already committed under docs/design/<slug>/):

1. Package on the way — the skill composes launch-design-handoff, which unpacks in scratch, names every canvas in the export by title, inventories the prototype, commits docs/design/<slug>/ (prototype verbatim plus handoff.md), and returns without sizing or routing. Its documenting questions and the build's blocking questions form one interview under the interaction contract.
2. Resolve and echo — the canvas (matched to the user's words, asked when ambiguous), the designed viewport ("1080p" = 1920×1080), the product mapping, the branch — echoed once before any code.
3. Write the contract first — docs/design/<slug>/fidelity.md: every reachable state (found by reading the prototype's script and driving it in a browser), the token/component/data maps, every row pending, committed before the build so nothing is narrowed quietly.
4. Build state by state from the prototype's own markup and styles (agent-browser get styles / get box on the running prototype where indirect), replacing the design it supersedes in the same change; real data or a recorded decision, never sample data in product code.
5. Verify side by side — two agent-browser sessions at the same viewport, the prototype from its committed .dc.html and the product from launchrail dev, each state captured on both sides and looked at, verdicts match / deviation / accepted (with a concrete reason) / unverified recorded per row. Captures under .launchrail/state/design/<slug>/, gitignored and regenerable.
6. Gate and close — verify --fast, full verify, launch-code-review, the record committed with the build, a PR of its own, evidence reported (matched / accepted / unverified) with the rail banner.

Changes

• launch-design-implement/SKILL.md (user-typed, disable-model-invocation) and fidelity.md — the protocol and record template, referenced by name from the other skills rather than paraphrased.
• launch-design-handoff — names the canvases in an export (one .dc.html per canvas; only the named one is the subject), inventories an interactive prototype with the protocol's state inventory, records the designed viewport, offers the faithful build route beside tweak / feature / redesign, and returns without routing when the build skill calls it to package a drop. It remains a command of its own only for the sized paths.
• launch (design-first paragraph, keywords), launch/workflow.md (the † user-typed list, a Returning from Claude Design paragraph), launch-ralph-implement (a ticket that names a package follows the protocol), launch-browser-smoke (the named design comparison is the protocol's side-by-side).
• ADR 2026-09-17-design-fidelity-build — amends 0018 (a second user-typed build command, for a different input; the ticket door is unchanged) and 0024, extends 0034; a two-invocation flow is a recorded rejected alternative; registry index regenerated with adr index.
• skills.test.ts pins the user-typed set at five; README and getting-started document the command; examples/hello-launchrail regenerated from the built CLI (also picks up the 1.17.0 version line, the heal-ADR-guidance migration entry, and the managed minting contract that the last release added).

Verification

• pnpm build and pnpm test: 20 files, 256 tests passing.
• node packages/cli/dist/index.js adr index regenerated the registry; doctor on the regenerated example reports healthy (17 checks, the usual package-manager warning).

🤖 Generated with Claude Code

https://claude.ai/code/session_01UCzoFemkzM6yMcHJifsa9P