fix(viewer): make the first camera fit uncancellable and frame to the real frustum

[w1ne]

Embedded models can display an empty viewport. Two independent defects in the same area, both fixed here.

1. A cancelled first fit is never re-issued

lastFitBounds.current is stamped when the fit is requested (line 58), but the fit is delivered over a ~600 ms lerp in useFrame — and that lerp is cancellable: onUserInteractStart nulls targetState on OrbitControls' start event. Once cancelled, every later run of the effect hits if (boundsStable) return under the 10 % tolerance, so the fit is never re-issued. The camera strands wherever the tween happened to be, permanently.

The default eye (40,40,40) sits right on the corner of a box(40,30,20) (which spans [0,40]×[0,30]×[0,20]), so an early abort leaves the camera at or inside the part — a blank frame.

Measured on the live embed, one click during the tween:

click	camera at click	camera at t=10 s
120 ms after mesh	(55.60, 52.72, 43.06)	(56.32, 53.31, 43.21) — never moves again
immediately	(45.50, 44.49, 41.08)	(46.77, 45.52, 41.33) — never moves again

On an embed this is very likely to fire: viewers scroll or tap the moment a page loads, and the damage is permanent rather than recoverable.

Fix: targetState gains immediate, set when lastFitBounds.current === null. A first framing lands in one frame and refuses cancellation; later re-fits keep the tween and keep yielding to the user. A pending pose is also held while controls is still null — retiring one before OrbitControls publishes itself lets its origin-initialised target re-aim the camera at (0,0,0).

2. The fit distance is wrong — no interaction needed

distance = radius * 2.8 is wrong twice: framing a sphere needs radius / sin(fov/2) (2.924 at fov 40), not the tangent (2.747); and camera.fov is the vertical angle, with no aspect term at all. So a portrait or narrow viewport overflows.

Live probe, untouched, 420×780: shipped distance 75.39, required 151.2 — the model overflowed all four edges. At 16:9 it needs 85.0 vs 75.4, which is the ~1° of clipping visible at the frame edges.

Fix: fitDistance(radius, fov, aspect, margin=1.08) using sin of the narrower of the two half-angles, used by all three fit paths (auto, home, focus).

How it was diagnosed

Not by reading. Playwright + real Chrome against production with window.__THREE_DEVTOOLS__ defined before load, wrapping renderer.render(scene, camera) to log the camera every frame. The page reported v0.15.0 (115a262) — exactly develop HEAD.

That trace disproved two plausible theories: there is no demand frameloop (frameloop/invalidate() appear nowhere in src/), and nothing else writes the Studio camera (CameraHandler.tsx is the only writer; demoPlayer/* is a separate viewer). With zero interaction the auto-fit does complete on its own by t≈3 s.

/p/<slug> shares the cause — embed.$slug.tsx → FunnelViewer → the same Viewer → the same CameraHandler.

Verification

• tsc --noEmit clean; eslint clean on all 3 changed files.
• npx vitest run src/studio → 100 files, 660 tests, all passing.
• Negative control (re-run independently): reverting both source files with the test kept fails 4 tests — expected -12.11 to be greater than 0, -1.18, -12.45, -1.11. The value is the signed slack of the tightest frustum plane in world units; negative means clipped. Restored → 6 passed.
• The two anti-twitch tests pass in both states by design: a param nudge must not wipe the user's orbit, and a genuine shape change must still re-fit with a grabbable tween. They prove the documented behaviour survives.

Assertions are on the camera pose the controller actually produces, driven through the real three-stdlib OrbitControls — not on pixels.

Not verifiable here

No local browser check: npx vite won't start on this machine — Cannot find module '../lightningcss.darwin-arm64.node', the same broken native binary behind the known viteReviewLiveEndpoint failure. It blocks npm run dev and vite build generally, not just this change. The live-renderer probing above was done against deployed production instead.