[Onlyti]

L5 per-wheel camber, toe and hardpoint-coordinate output on the strut path with a corner model, and one VehicleParams.rack_per_rad key in place of the two rack-per-radian constants (work order 48c, 0.8.2). Dynamics unchanged: without the new key and without reading the new outputs every level is bit-identical to 645780c.

What changes

• SimOutput: wheel_camber, wheel_toe [rad] (equal to free_3d_wheel_camber/toe, product toe of theory 14 §14.16), wheel_kin_valid, hardpoints (per wheel, count points [m]). L1-L4 and wheels without a corner model: 0, valid false, no points.
• Frame (PO verdict D1 + origin verdict, option 1): vehicle design frame, x forward, y left, z up, origin on the ground below the front-axle centre. Rear part YAML shifted by axle_offset_x = -wheelbase - w_rear.x; right wheels are the left part mirrored in y. wheel_center_dy = w.y - track/2 is recorded, coordinates are not moved in y.
• free_3d_hardpoint_layout() (names, axle_offset_x, wheel_center_dy, excluded), free_3d_wheel_kin_valid(), free_3d_wheel_hardpoints(), free_3d_rack_dy(); all exposed to Python.
• VehicleParams.rack_per_rad (default 0.08 m/rad), used by L4 and L5 roadwheel mode; YAML / Python / to_yaml round trip, invalid values raise.
• Python: vdsim.attach_l5_hardpoints(...); vdsim_lab.attach_kinematics delegates to it on L5. The L3/L4 attach keeps returning False on L5.
• .vdtrace schema stays 0.5: opt-in wheel_camber, wheel_toe (4,) and hp_<wheel>_<key> (3,) channels (regex ^hp_(fl|fr|rl|rr)_[a-z0-9_]+$), manifest hardpoints block; an invalid wheel is NaN.
• Five-link corner (PO K4 verdict, option 1): points not exported (count 0, layout names empty, excluded reason; manifest points: {} + excluded). Camber and toe are still reported (they are the values the tyre uses).

Gates (48c sec. 4, fixed in 5ec3702 before the implementation)

ID	Result (local, HEAD 4b47562)
K1 bit identity vs 645780c	24/24 serialized runs identical (L1-L4, L3/L4 with native kinematics, L5 penalty / strut / three corner-model pairs; generic_sedan and fs_aero_example). RL pedal and accel obs+reward digests identical. Closed-loop deterministic arrays identical (only wall-clock t_compute_s / compute_* differ)
K2 camber/toe = free functions every tick	pass, both pairs
K3 static coordinates, attach and after reset()	pass, 1e-9 m
K4 link lengths, bump +-30 mm x rack +-22.5 mm	MacPherson, trailing arm, double wishbone pass (prescribed sweep and every session tick); five-link: see below
K5 mirror	pass
K6 rack_per_rad	0.08 given = absent bit-identical (L4, L5); 0.04 gives exactly half the rack
K7 negative	per-tick hardpoint refresh removed -> K4 session test fails; restored -> passes
K8 Python = C++ bit-identical, L5 attach	l5_hardpoints ctest passes
K9 RL preset throughput, A/B 5 rounds alternating, cores 0-9,12-21	median change within the 3 % line, no opt-out switch needed. L5 corner-model single session cost recorded as information

C++ and Python counted separately: C++ 16 new tests (unit 4, integration 12), Python 1 new ctest.

Changed after measurement

• Five-link (PO verdict on the K4 halt, option 1; line unchanged): the five-link driven pose is a least-squares fit with 7 residuals for 6 unknowns, so its links do not hold their lengths to the K4 line. Implementation commit 51cad6d stops exporting its points; test commit c895d4f judges K4 on the three other topologies, asserts the exclusion (FiveLinkCornerExportsNoPoints, Python manifest check) and keeps the five-link sweep runnable as L5HardpointKnownFailure.K4FiveLinkPrescribedSweep (label known_failure, excluded from the validation preset). 5ec3702 is untouched.
• K5 on the session output skips an axle that exports no points; the prescribed K5 on the five-link corner model still runs.

Documentation impact (W1)

CHANGELOG [0.8.2] (Added, Known limitations: L4 no output, five-link exclusion, wheel_center_dy), docs/theory/14_hardpoint_kinematics.md §14.17 (frame, point lists, rack_per_rad, trace channels and manifest), docs/EXPERIMENT_API.md trace channel note, docs/VALIDATION.md currency PENDING. Doc audit: FAIL set on HEAD equals 645780c (pre-existing A1 only).

Shared files with work order 47

free_3d_dynamics.cpp, params.*, bindings.cpp: whichever lands second merges main (no rebase).

Local verification

ctest --preset validation --parallel 4: all passed (main 595 + 17 new). ctest -L known_failure: the two R2 items and the five-link K4 item, as labelled.

CI measurement

PENDING: same-head complete successes x3, currency confirmation, final run and wheels. No merge, tag or publication performed.

🤖 Generated with Claude Code