investigate(springbone): inside-collider diagnostic + substep ordering

[arkavo-com]

Summary

• Adds VMK#237 inside-collider diagnostic harness (per-bone last-write buffer the sphere/capsule kernels populate with (shape_type, angleLimit, distance, boundary, penetration, group_matched), opt-in via renderer.setInsideColliderDiagnosticsEnabled(true)). Sentinel shape_type values distinguish "never ran" / "kernel-entered no match" / "outside-branch with inside-flag=0" / "inside-sphere fired" / "inside-capsule fired" so future investigations can pin parameter-propagation bugs without re-instrumenting.
• Substep ordering: sphere + capsule collision now interleave with the distance constraint iteration loop (SpringBoneComputeSystem.swift's XPBD loop). Inside-* containment is re-applied after every distance pass instead of only once at the end; plane stays outside the loop.
• New regression test testInsideSphereAngleLimitReachesShader documents that BoneParams.angleLimit reaches the shader correctly (1.047 rad / 60° read by every non-root bone).

Finding

The conformance bucket collapse for springbone_extended_isphere_* variants is not a VMK propagation bug. The diagnostic confirms parser / GPU buffer build / shader branch / group filter / per-joint angleLimit are all wired correctly. Chain joints stay at distance 0.0046 / 0.0506 / 0.1005 m from the sphere centre, well inside the 0.20 m and 0.40 m boundaries at pmed / ploose, so inside-sphere collision correctly never engages; only ptight (r=0.10) is tight enough to clip the chain. Angle-limit values 30°/60°/90° collapse because the chain's actual swing never exceeds 30° from bind.

vrm-conformance suite addressed it on their side by reducing inside-sphere/capsule placement radii from [0.10, 0.20, 0.40] to [0.04, 0.06, 0.08]. Bucket count 7 → 11 (+57%), non-zero penetration across all 18 swing variants.

Behaviour change

Sphere and capsule collision dispatches now run once per distance-constraint iteration instead of once at the end of the substep. For models that previously sat inside their colliders without distance constraint forcing them out, behaviour is unchanged. For models where distance constraint can pull a joint back outside a containment surface, the joint is now pushed back in within the same substep instead of carrying the violation forward. Plane behaviour unchanged.

Performance: constraintIterations=4, substeps=4 → 16 dispatches/frame for sphere+capsule combined vs. 4 before. Each dispatch is one bone-count grid (≪ 100 threads typical), trivially cheap.

Test plan

• swift test --filter SpringBone --disable-sandbox — 170 tests, 0 unexpected failures (1 expected failure caught by XCTExpectFailure)
• swift test --filter SpringBoneExtendedColliderBehaviorTests --disable-sandbox — 7 tests, including the new diagnostic test
• swift build clean
• make shaders-macos clean
• AvatarSample_A.png regenerated, no visible artifact (88-byte delta from prior render)

Closes #237

🤖 Generated with Claude Code