[KunalSiyag]

What

occupants.sync re-ran the seat IK every frame for every seated driver and rider in range, on poses that cannot change: a car driver's sit pose is a pure function of the car's basis (the body rides the car rigidly and the pedal/wheel markers are car-space), and a rider's legs are a pure function of the crank angle, their arms of the basis alone.

The seat points, driver markers and limb bone chains were already cached; this memoizes the IK itself:

• Car drivers re-run _driver_limbs once at seating and then only when the car's global_transform.basis changes (a grade, or a turn while vis_yaw catches up).
• Riders re-run the leg IK only when the crank angle or the bike's basis changes, the arm reach only on a basis change, and the clip seek only when the legs re-ran. A stationary crank now early-outs after the memo update (arms at rest are the correct final pose; previously the early-out preceded the memo write, leaving the memo key stale).

Pin

New public counter ik_runs (test-only read). fp_traffic_test._check_driver_ik_memo asserts a held car adds no runs across repeat syncs and a turned car adds exactly the re-run — the two properties a regression of either memo would break.

Two bugs the suite caught in the first cut (fixed here)

• The rider memo lives at entry slot 5, which _place read as the car-only sink float: float(Dictionary) errored every frame and rider roots froze off their seats. The sink read is now gated to car drivers.
• The check itself leaked state: _look_at_car leaves the hero beside a moving car, which deterministically slowed the later resident-driver check by 23 s of sim time (traffic yielding to the hero). It now restores the hero position and the rotated car on exit.

Validation

• fp_traffic_test PASS (incl. the three new ik-memo pins; resident driver 27.2 s vs main's 26.8 s).
• tools/gate.sh quick: 180/181 steps PASS; the single FAIL was fp_car_doors_test on the Male_Adult_12_head.tres "possible cyclic resource inclusion" engine ERROR — a load-race flake also seen on other branches' gate runs this week; the suite's own verdict in that run was PASS and a standalone re-run is PASS with zero such errors.
• pre-commit (gdlint/gdformat/complexity/repo-checks) green; api docs regenerated (ik_runs).

[github-actions]

Automated review

No new findings on the 3 changed files (gdlint, complexity, TODO policy, dead code).

Run python3 tools/checks/run_all.py locally for the full set of checks.