Epic: Romero-style performance roadmap (10 prioritized optimizations)

## Source
Triaged from an external performance review that applies the id Software ("Romero / Carmack") cultural and engineering lens to VRMMetalKit. The roadmap is mirrored in `docs/PERF_ROADMAP.md` on branch `perf/romero-roadmap` for in-repo discoverability.

## Principle
> *"You don't ship a fast renderer. You ship a renderer, measure it, and then for the next six months you delete code until it's fast."*

## Prerequisite — must land first
- [ ] #156 — Benchmark CI regression gate. Romero rule #1: never optimise anything you haven't measured. Every item below claims an impact; none of those claims are real until we have a baseline.

## Items (priority order from report)

| # | Romero item | Issue | Status |
|---|---|---|---|
| 1 | Publish baselines | #156 | already filed |
| 2 | Outline pass merge (tile memory / instanced draws) | #192 | new |
| 3 | SpringBone sleep gate when avatar is at rest | #149 | already filed |
| 4 | MToon pipeline specialisation via `[[function_constant]]` | #193 | new |
| 5 | Mask-dispatch morph targets to face-region vertices | #194 | new (orthogonal to #150) |
| 6 | Split VRMVertex into position-only + attributes | #195 | new |
| 7 | Demote MToon fragment intermediates to `half` | #196 | new |
| 8 | Tile-memory shared render pass | folded into #192 | — |
| 9 | MPS Kalman for ARKit smoothing | #198 | new |
| 10 | Dual-quaternion joint palette | #197 | new |
| + | GPU occlusion queries for crowd avatars | #199 | new (extends #91 / #154) |

## Expected impact (from report — claims, not measurements)
- First three items: ~30–40 % frame-time reduction on a single-avatar scene.
- Full list: roughly 2× on crowd scenes where the priority system and occlusion culling compound.
- **All numbers are conjecture until #156 produces a baseline.**

## Cultural commitments the report calls out
- Every PR runs against a fixed reference scene (`AvatarSample_A_1.0.vrm.glb` is already in the repo).
- No performance regression merges without explicit sign-off.
- Publish a budget ("one avatar at 120 Hz on iPhone 15 Pro under 4 ms GPU" or similar) and defend it.

## Cross-references to existing in-flight work
- #88 — outline via ICBs (alternative architecture to #192's tile-memory merge; pick one)
- #91 — crowd scalability (umbrella for #199 + parts of #192)
- #124 — skinning dedup (prerequisite for #195 + #197)
- #125 — per-light dedup (prerequisite for #193 + #196)
- #154 — frustum culling refinement (precondition for #199)
- #155 — MToon maintenance (intersects #193 + #196)
- #150 — morph "skip when unchanged" (orthogonal to #194)

Start by reading docs/PERF_ROADMAP.md on branch perf/romero-roadmap and issue #156, which is the stated prerequisite for measuring the listed optimizations. Review the dependencies and alternatives in issues #192, #193, #195, #197, and #199, along with the fixed AvatarSample_A_1.0.vrm.glb reference scene. Done requires turning this broad roadmap into measured, separately actionable performance work rather than implementing one obvious change.

Written by the indexing model from the issue text.