feat(adapter): make parity-lit — contract + visual parity (WHYNOT-WP-0002 T08)

Render every <wn-*> in a real browser (Playwright, --no-sandbox; reuses an
external static server when present) and write the adapter-contract parity result
to adapters/lit/parity/_parity.json:

- Contract parity: element must upgrade + carry no attribute-mismatch vs IR
  (computed statically via scaffold.componentDrift, avoiding runtime
  type-coercion false positives). prop-missing is a coverage note, not a failure.
- Visual parity: render smoke (non-empty + positive box) + per-component
  screenshot artifact (gitignored). Pixel-exact regression stays with the
  Playwright baseline suite; IR exemplars are gallery cards, not single-component
  baselines, so they are the human reference, not an auto pixel gate.
- Result: 10 components, contractFail=0 visualFail=0, PipelineStrip skipped
  (wn-pipeline-strip rename drift). Exit 4 on failure.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
2026-06-30 09:16:13 +02:00
parent 552d8fe926
commit e4e3fe069c
5 changed files with 327 additions and 2 deletions

View File

@@ -13,8 +13,35 @@ Per the contract, an adapter is **scaffold + drift-detect**, never a rewrite:
| Concern | Behaviour | Status |
|---|---|---|
| **Tokens** | **Fully generated** from `ir/tokens.json` into the `:root` block of `src/styles/colors_and_type.css`, between `@generated tokens` markers. Deterministic — re-running with an unchanged IR is a byte-identical no-op. The hand-authored type/utility CSS after the block is preserved. | **done (T06)** |
| **New component** | Generate a `<wn-*>` Lit stub from the IR contract's prop→attribute map + a behaviour `TODO`. | T07 |
| **Changed component** | Emit a **drift report** (`adapters/lit/drift/<Name>.md`) — never overwrite the hand-authored element. | T07 |
| **New component** | Generate a `<wn-*>` Lit stub (`adapters/lit/stubs/<Name>.js`) from the IR contract's prop→attribute map + a behaviour `TODO`. **Write-once** — into a staging dir, never the hand-authored tree; the human integrates it. | **done (T07)** |
| **Changed component** | Emit a **drift report** (`adapters/lit/drift/<Name>.md` + machine `_report.json`) — never overwrite the hand-authored element. | **done (T07)** |
### Drift severity
`make adapt-lit` exits `3` only on **actionable** drift — `prop-missing`,
`attribute-mismatch`, `variant-axis-missing`, `tag-mismatch`. **Informational**
issues do not gate: `non-portable` (React `style`/callbacks that inherently have
no attribute form — the Lit element is right to omit them) and `prop-extra` (the
Lit element is richer than the minimal React designbook). Resolve actionable drift
per `.claude/rules/designbook-propagation.md` (fix the stack, or change the language
in Claude Design and re-propagate — never a stack→React back-edit).
## Parity — `make parity-lit`
`adapters/lit/parity.mjs` renders every `<wn-*>` in a real browser (Playwright) and
writes `adapters/lit/parity/_parity.json` (the contract's parity-result shape):
- **Contract parity** — each element must upgrade and carry no `attribute-mismatch`
vs its IR contract (computed statically, so no runtime type-coercion false
positives). A prop the element *lacks* is a coverage note (already surfaced as
drift), not a parity failure.
- **Visual parity** — a render smoke: the element renders non-empty with a positive
box; a screenshot is saved to `adapters/lit/parity/<Name>.png` (gitignored) as the
artifact. The `ir/exemplars/<Name>.html` are designbook **gallery cards** (a grid
of all variants), not single-component baselines, so an automated pixel diff
against them is not meaningful — per-component Lit appearance regression is owned
by the Playwright baseline suite (`tests/visual/`); the exemplar is the human
visual reference. Exit `4` on a contract or render failure.
## Directionality