Cloth-fidelity post-merge cleanups (from final review)

Non-blocking follow-ups identified during the final whole-branch review of cloth-collision-fidelity (fix/stagger-collision). Terse list, one issue so they don't scatter:

- **Oracle build caching across frames**: `SkinMeshOracle.build(model:)` is rebuilt 375 times per gate run (~15 min of suite cost); the underlying skinned mesh doesn't change per query within a pass — cache/reuse across frames where safe.
- **`humanoidBone(ofNode:)` map hoist**: same shape of cost as the above — a map built once and reused instead of recomputed per lookup.
- **Bust chains skin the body surface**: the flag perturbs oracle ground truth between flag-off/flag-on passes because Bust spring joints skin body-surface geometry; freeze Bust during oracle build so the ground truth doesn't shift between the two passes being compared.
- **Dual-sabotage repeatability**: the two independent sabotage halves (measurement-stubbed, segment-disabled) in the oracle gate should be re-runnable via env-gated toggles rather than one-off manual scratch builds.
- **`dominantVertexCount` post-filter semantics**: needs a clearer contract for what counts as "dominant" when weights are close/tied.
- **Grid tie-break epsilon below Float ulp**: `SkinMeshOracle.SpatialGrid`'s `1e-12` tie-break constant is below `Float`'s ulp at the magnitudes involved and should be revisited.
- **Large-ring diagonal path untested**: the diagonal-query gap noted in `SpatialGrid.nearest`'s comments (ring growth beyond `fallbackRingThreshold` via the Euclidean-vs-Chebyshev distinction) has no dedicated test.
- **Excluded-inventory vertex-count pins**: `BodySurfacePredicateTests`' excluded-material assertions don't pin vertex counts the way the included set does — drift there is currently silent.
- **Dead asserts and nil==nil grid queries** in `SkinMeshOracleMathTests` — cleanup pass needed.

Reference #381.

Start by reading SkinMeshOracle.build(model:), humanoidBone(ofNode:), SpatialGrid.nearest, BodySurfacePredicateTests, and SkinMeshOracleMathTests, then identify which follow-ups can be isolated. The work is done when the listed caching, Bust/oracle, repeatability, semantics, epsilon, coverage, inventory-pin, and assertion cleanups are implemented and the relevant tests cover them.

Written by the indexing model from the issue text.