feat(shell): re-apply VR reference-space offset on SceneRack hops (#283) - #284

Closes #283

Summary

Extends the per-scene VR spawn behavior from #263 to also fire on in-session SceneRack hops, not just on sessionstart. Pre-diff, hopping quadrics → tangent-planes mid-VR-session stranded the user at world Z ≈ +1.5 while the new plinth sat at Z = -2.125 — a measured ~3.6 m gap instead of the intended ~1.45 m (Brad's #282 smoke verdict). Lifts applyVRSpawnOffsetForExhibit out of the VR-branch closure into a mode-independent let slot the VR branch populates; switchExhibitNow calls it after the mount + per-scene rack/camera writes. Stacking-safe by construction — the helper always derives from the unoffset baseXRReferenceSpace snapshot, so N consecutive hops collapse to a single per-scene offset rather than a sum.

Test plan

Automated:

• npx tsc --noEmit clean
• npm run lint clean
• npm test — 604/604 pass
• /spar (Sonnet oppositional review): PUSH, zero findings — stacking, boot-mount isPresenting no-op, pancake-mode null hook, sessionstart double-application, disposers, and HMR-generation scoping all verified.

Headset smoke on the Cloudflare PR preview (Quest 3S):

• Enter VR on quadrics — user-to-plinth-front distance reads ~1.45 m (regression check vs. feat(staging): per-scene plinth + spawn derivation from cutout (#263) #282).
• In-session hop quadrics → tangent-planes — distance reads ~1.45 m to the new plinth (not ~3.6 m).
• In-session hop quadrics → gradient-levels — ~1.45 m.
• In-session hop quadrics → saddle-extrema — ~1.45 m.
• Reverse hops (tangent-planes → quadrics, gradient-levels → quadrics, saddle-extrema → quadrics) — ~1.45 m to the new plinth.
• Multi-hop chain (e.g. quadrics → tangent-planes → gradient-levels → saddle-extrema → quadrics) — no offset stacking; each step lands at ~1.45 m.
• Mid-VR-teleport read: the rack + plinth already rearrange under the user on every hop today (shipped in feat(staging): per-scene plinth + spawn derivation from cutout (#263) #282); confirm bringing the user along keeps the rearrangement consistent rather than disorienting.

Pancake regression check (preview URL or npm run dev):

• Pancake SceneRack hops continue to reposition the camera per-scene with no VR-path interference (call site is ?. no-op when applyVRSpawnOffsetForExhibit is null).

🤖 Generated with Claude Code