fix(babel-plugin): stamp only tags that are actually DOM

[divshekhar]

The most severe defect in the 2026-09-08 field export. Extracted unchanged from eb3e6c57 on fix/pr-list-tells-the-truth, where it rode along inside an onboarding commit under a PR titled about the open-PR list (#872) — invisible to any reviewer, and lost if that branch were reworked. See my comment on #872 for the suggested split.

The defect

"Lowercase means host element" is true only for React DOM. React is a reconciler interface: react-three-fiber, react-pdf and ink define their own lowercase intrinsics, and those host instances are not nodes and do not take attributes.

R3F is the case that bit. Its applyProps treats any prop containing a dash as a pierced property path, so data-reticle-source is walked as data → reticle → source. There is no data object on a three.js instance, so it throws:

Uncaught Error: R3F: Cannot set "data-reticle-source". Ensure it is an object before setting "reticle-source".
    at applyProps (react-three-fiber.esm:434:79)
    at commitUpdate (react-three-fiber.esm:8631:5)

The throw is inside R3F's commit phase and unhandled, so it does not degrade the viewport — it unmounts the entire React app to a white screen.

Why it is severe rather than cosmetic

• The option that causes it is on by default.
• It fires on an element update, not a mount. In the field the app imported a STEP file, ran a volume extraction and rendered geometry for ~20 minutes across ~15 passing verdicts before a newly generated mesh took it down mid-pipeline, losing all in-memory setup state.
• So reticle init looks clean, a session connects, verdicts pass, and the app dies much later — which reads as an application bug. The reporter initially diagnosed it as one and only found us by reading the stack trace.
• Its documentation names the React version as the risk axis. The actual risk axis is whether the tree contains non-DOM reconciler elements.

Net: an entire application category — CAD, simulation, 3D configurators, map and model viewers — is not merely unverifiable but actively broken by installing us.

The approach

An allowlist, not a heuristic. The set of custom-reconciler intrinsics is unbounded and cannot be enumerated; the set of HTML and SVG tags can. Wrong conservatively costs one source pointer on an exotic element. Wrong the other way costs the user their entire app.

Stated rather than implied in dom-tags.ts: a few three.js class names collide with real DOM tags — line is both <line> in SVG and THREE.Line, and so is audio. A tag name alone cannot separate them, and a lexical "is it under a <Canvas>" walk only sees the file being transformed. The complete answer for a three.js app remains reticle({ sourceMapping: false }), which init now writes on its own when it finds react-three-fiber in the manifest.

Verification

pnpm format:check, pnpm lint, pnpm typecheck green; 15/15 babel-plugin tests.

Not claimed: I have not driven a real react-three-fiber app against this build. The unit tests pin that the listed non-DOM intrinsics are skipped and DOM tags still stamp; the end-to-end proof that the white screen is gone is the thing #891's awkward-app corpus exists to make routine.

🤖 Generated with Claude Code