Fix trajectory shake: smooth IK outliers, fix sim dt jitter

[Jepson2k]

Summary

Three independent fixes that together eliminate visible robot shake during the precision sweep section of the demo, plus an @njit rewrite of the IK-outlier smoother with regression tests.

Why

Recording multicast status during precision.py playback showed real-position discontinuities of ~4° per 20ms status frame on J4/J6 during the RX/RY/RZ wrist sweeps. After ruling out controller pacing, sim drift, and downstream rendering, the root cause was traced to three independent issues:

1. pinokin LM IK has no continuity preference at wrist singularities. Near J5 ≈ 0, J4 and J6 axes coincide; the IK has a one-parameter family of valid solutions and the LM solver can pick one several degrees off the natural chain. FK still matches, but TOPP-RA's time parameterization treats the off-branch sample as a steep joint motion and cranks up local velocity.
2. Wallclock dt in the simulator inherits the host scheduler's jitter. On slow ticks the simulator under-advances Position_in; on fast follow-up ticks it over-advances. Visible as ghost velocity spikes in recorded status even when the commanded trajectory is bit-identical across runs.
3. PATH_SAMPLES default of 50 produces ~7mm cartesian spacing on a 350mm linear move — sparse enough that the IK-chain step size varies enough between samples to confuse threshold-based outlier detection.

Implementation

1. parol6/motion/trajectory.py — _smooth_singularity_outliers runs after batch_ik succeeds:

   • Detection is median-relative: a chain step is an outlier if its magnitude exceeds _IK_OUTLIER_RATIO (10×) the chain's own median step. No absolute angle threshold; no separate bookend/path-length test. The ratio knob is well-behaved because real LM hops exceed 20× while normal motion stays within ~5×, and the rule is invariant to PATH_SAMPLES, speed, and move type.
   • For each contiguous run of outlier samples, linearly interpolate over the run plus _IK_OUTLIER_PADDING (4) samples on each side. Padding absorbs the LM seed-bleed that contaminates a few samples after the hop. FK deviation is sub-mm since both bookends lie on the cartesian path.
   • Implemented as @njit(cache=True) with explicit scalar loops; only allocations are one np.empty(n-1) for step magnitudes and whatever np.median does internally — both bounded by path length, called once per plan, not per tick.
   • Pre-compiled in parol6/utils/warmup.py so the JIT cost is paid at server startup, not on first plan.

2. parol6/server/transports/mock_serial_transport.py — pass cfg.INTERVAL_S to the JIT motion simulator instead of wallclock dt. Sim now advances by exactly one control tick per call, matching firmware behavior.

3. parol6/config.py — bump PATH_SAMPLES default 50 → 200. Denser sampling makes the median-step estimate robust on short moves. PAROL6_PATH_SAMPLES env override unchanged.

4. examples/precision.py — sync to match the current demo_showcase.py precision section.

5. tests/unit/test_motion.py — new TestSmoothSingularityOutliers class covers: smooth-path pass-through, single-sample hop (3 bad samples + 2·pad patched), wide rising shelf (one contiguous run), padding clamping at array start/end, n < 3 no-op, zero-motion no-op (median=0), sub-threshold and at-threshold non-trigger, and a randomized fuzz that verifies positions[0] and positions[-1] are always preserved.

Test plan

• pre-commit run --all-files clean
• pytest clean (231 passed, 0 failed)
• Manual: precision.py runs without visible J4/J6 shake during RX/RY/RZ sweeps
• Recording multicast status during a full demo: peak per-frame J4 velocity stays below the 2.45 rad/s limit (was peaking at 4+ rad/s before)

Notes

• _IK_OUTLIER_RATIO is insensitive across [10, 20+]; chosen at 10× as a safe lower bound.
• The smoother only runs on the all_valid branch of JointPath.from_poses — partial paths bypass it (no behavior change in failure cases).
• pinokin pin to v0.1.6 was already on main (abc765f) and is inherited here.