fix: foot-IK idle leg buzz (wrong leg orientation targets)

[blugart-dev]

What

Fixes the idle leg vibration in foot IK — a latent 0.3.0 quality bug surfaced while validating the (now-parked) 0.4.0 stumble work. Substrate hardening; lands on main before 0.4.0 resumes.

Root cause (data-backed)

The two-bone IK solver built each leg segment's orientation target from scratch with an assumed local-axis convention (_basis_looking_along → Basis(s, -d, f)). That convention sits ~130° off the real Mixamo bone bases. The velocity spring then perpetually tried to twist the legs to those joint-unreachable orientations → sustained idle buzz.

Measured headlessly on the real ybot asset at rest (target error = angle between the IK target basis and the actual bone; ang vel = idle angular velocity):

Bone	target error before → after	idle ang. vel before → after
UpperLeg	130° → 0.6°	1.1 → 0.17 rad/s
LowerLeg	33° → 1.0°	1.8 → 0.52 rad/s
Foot	11° → 2.4°	1.7 → 1.17 rad/s

(The investigation also confirmed the velocity-spring core is sound — it tracks consistent targets calmly and is the same formula Jolt's own DriveToPoseUsingKinematics uses. It was faithfully chasing garbage targets, not malfunctioning.)

The fix

Express each leg segment's orientation as a swing of its animation basis onto the new IK bone direction — the same delta-from-animation approach the foot already used for slope correction (and why the foot's error was smallest). When the IK direction equals the animation direction (no adjustment needed), the swing is identity, so the target equals the animation pose → zero error.

• Convention-agnostic: no longer assumes a Mixamo local axis, so non-Mixamo rigs benefit too (aligns with the role-based generality work).
• Positions unchanged — knee, foot target, and pelvis drop are untouched; only segment orientations are corrected. Foot planting / pelvis / leg-length behavior is identical.
• Removes _basis_looking_along; adds a degeneracy-hardened _swing helper.

Validation

• Headless measurement above (target error 130° → <2.5°; leg buzz ~collapsed).
• GUT 113/113 — all existing foot-IK tests (planting, pelvis-never-lifts, leg lengths, role-mapped non-Mixamo) still pass.
• Visual in-editor confirmation pending (idle legs should be calm).

Note

The remaining ~1 rad/s residual on the foot is the spring's steady-state floor — addressed by the separate spring-hardening PR (settle deadband + solver-cooperative velocity), the second half of the substrate hardening.

🤖 Generated with Claude Code