VDSim 핸드오프

작성: 2026-06-26 · main @ bfcc5cd · 391/391 ctest green · CI run 28213027035 green

Structured review (2026-06-21): portable fixes on main worktree — see docs/evidence/review/REVIEW_2026-06-21.md. GUI v3 frozen on VDSim-Thesis only (a44b593). VLA plant API on main via cherry-pick.

0. 2026-06-26 launch pre-flight — DONE (publish HOLD)

A cherry-pick (thesis → main, GUI excluded): 13b335c wheels CI · 6360aee quickstart · 5eaee08 demo · a54447a VALIDATION · 024dc4b plant (+ d0dbf46 drive_split_front, 97c3d7a Win32 plugin loader, bfcc5cd wheel vehicles/ + vdsim_plant.py).
C CI: workflow_dispatch run 28213027035 — wheels cp310/311/312 × linux+win + sdist + verify green; publish skipped (no tag).
Local conda (Linux): clean py3.10/3.11/3.12 — import + Sim 3종 (sedan L2, 두 번째 차량 프리셋 L2, L1) + vdsim-quickstart + demo_grip_loss.py green on CI artifacts (0.7.0.dev26).
D setuptools-scm: 0.7.0.dev26 now; SETUPTOOLS_SCM_PRETEND_VERSION=0.7.0 → 0.7.0 (> 0.0.0).
HOLD: no v0.7.0 tag / PyPI / public release until PO GO.

0. 2026-06-25 launch pre-flight P0 on main — DONE (publish HOLD)

Wheel CI (P0-1): .github/workflows/wheels.yml — cibuildwheel + verify_wheel_packaging.py.
HOLD: no v0.7.0 tag / PyPI publish until PO GO.
Plant: python/vdsim_plant.py + test_vla_plant.py cherry-picked from thesis (4f948a5).
P0-2 quickstart: examples/quickstart.py + vdsim-quickstart; README ~6 s cold (measured).
P0-3 demo: examples/demo_grip_loss.py → docs/assets/demo_grip_loss.gif (headless, pillow).
P0-4 VALIDATION: currency v0.5.1+, 391/391 ctest on main (404 on thesis with GUI); README validation summary; tire cross-check wording (not product parity).

0. 2026-06-23 CmdL1 per-wheel torque binding fix — DONE

Bug: def_readwrite on std::array<double,4> returned a copy — cmd.motor_torque[i]=x was silent no-op.
Fix: MutableWheelArray in-place view + vdsim_plant._fx_to_cmdl1 whole-array assign.
Tests: test_cmdl1_binding.py, test_vla_plant.py.

0. 2026-06-20 core quality + T3 VLOW — DONE

Shipped: - SessionKind — Standard / DirectControl on SimConfig (replaces direct_control_path) - KC compliance — bushing toe/camber under lateral tire load in L3/L4 hard-joint step - T3 VLOW — TireParams.low_speed_mode + vlow_speed_eps; tests TireVlow.* - Chrono KC — gen_kc_reference fills reference CSV; ChronoKcParity active (rebound reach ~−14 mm)

Verify: cmake --build build -j && cd build && ctest --output-on-failure

0. 2026-06-20 thesis → main core port (partial cherry-pick) — DONE

Tagged: vdsim-thesis-beta @ 6623f55 — VLA plant beta delivery freeze on VDSim-Thesis.

Cherry-picked to main: c4b5c0a (+ fixup) — core only, no python/vdsim_plant.py / VLA integration tests. - TireSetup per-axle/per-wheel + catalog/cosim/pybind + PerAxle/PerWheel tests - IVehicleDynamics::tire(wheel), make_direct_control_session, FrictionMapConfig - steer_deadtime via ActuatorModel; ladder_lowering.hpp; SimOutput wheel GT fields - VehicleParams.plant_path (default false; required by L2 plant-path branch)

Still thesis-only: VLA plant wrapper (24 commits + e5fb2c6), drive_split_front (#6 deferred).

Verify: cmake --build build -j && cd build && ctest --output-on-failure

0. 2026-06-25 Product freeze boundary — ACTIVE

GUI v3: preserved @ P0 parity gap 0 (commit feat(gui): v3 preserve).
FROZEN — no further GUI polish/features until ≥2 scenario-authoring segment adoption failures cite GUI absence.

Plant API (P1 wedge): open — thesis / external MPC (vdsim_plant, CmdL1 path).

Wheel CI (P0-1): .github/workflows/wheels.yml — tag v* → cibuildwheel + PyPI (trusted publisher).
HOLD: no v0.7.0 tag / PyPI publish / main merge until PO decision-4 gate (PO signal).
Pre-publish: python3 tools/verify_wheel_packaging.py wheelhouse/*.whl (no .tir, no withheld names or run-output folders, generic presets present; also takes the sdist).

P0-2 quickstart: examples/quickstart.py + vdsim-quickstart; README cites ~6 s cold (measured, not "5 min").
P0-3 demo: examples/demo_grip_loss.py → docs/assets/demo_grip_loss.gif (headless, pillow).
P0-4 VALIDATION: currency v0.5.1+, 404/404 ctest; README validation summary; tire cross-check wording (not product parity). ISO numbers unchanged (IsoBaseline gated).

0. 2026-06-23 CmdL1 per-wheel torque binding fix — DONE

Bug: def_readwrite on std::array<double,4> returned a copy — cmd.motor_torque[i]=x was silent no-op.
Fix: MutableWheelArray in-place view (__getitem__/__setitem__) + vdsim_plant._fx_to_cmdl1 whole-array assign.
Tests: test_cmdl1_element_torque_assign, test_longitudinal_fx_accel in test_vla_plant.py.

0. 2026-06-23 CO-00 scenario template gallery — DONE

Templates: Empty · Figure-8 · Straight · Skidpad (R=40 circle)
API: POST /api/setup/template · Scenario inspector 2×2 chip gallery
E2E: template chip button.tpl locator (accessible name includes hint text)

0. 2026-06-23 GUI v3 sim time + RN-02 + SH-05 — DONE

Sim time: stream t = plant timestamp − first frame (not wall epoch)
RN-02: Run Autopilot/Manual + manbar (↑↓←→ + touch) → /api/fleet + /api/manual
SH-05: HelpLink → mkdocs theory pages (Scenario road/path, Run pose/tire)

0. 2026-06-22 GUI v3 Run/Scenario polish — DONE

RN-06: per-wheel force triad (Fx/Fy/Fz arrows) in RunViewport · Forces toggle
Run telemetry: sim time t · per-wheel Fz/Fx/Fy/κ/α table
CO-06: waypoint summary table (downsampled) in Scenario inspector
Docs: concept art README — require live capture pairing for thesis figures

0. 2026-06-22 GUI v3 visual P0 — DONE

P0 fixes (vs concept art review): - Scenario map: path from setup.path_pts (no 0-point race) · grid axis labels (m) · waypoint # in edit mode - Run idle: preview vehicle mesh + path at fleet spawn (RunViewport setup_mode) - Play → Run navigation hardened · capture flow waits sync+start · 06-run-sim-*.png

Review: docs/evidence/review/GUI_V3_VISUAL_REVIEW.md · captures docs/evidence/review/gui-captures/

Regenerate: python3 tools/gui_v3_capture.py · eval: python3 tools/gui_v3_eval_bundle.py

0. 2026-06-22 GUI v3 scaffold + Scenario API — DONE

Concept art: docs/design/assets/ (6 PNG + README).

Scaffold: gui/v3/ — Svelte 5 + Vite · #/scenario default.

Scenario mode (wired): GET/POST /api/setup, /api/path, /api/fleet (load scenario), /api/scenario/save, map canvas (path + fleet drag), road μ/grade/bank, path preset, ▶ Play → /api/run/start.

Still placeholder: — (L4 hardpoint editor shipped in Build mode).

Scenario path edit (CO-07): map toolbar Edit path · drag ● · click add · Del · Esc.

Analyze mode (v3.0-beta): /api/compare · ISO maneuvers · yaw-rate overlay · Δ% table · CSV export.

Build mode (v3.0-beta): assembly · blueprint Export/Register · Parts library · Δ stats · linkage SVG · L4 hardpoint editor (pick/drag y–z, K&C preview, save to gui_custom).

Run mode (v3.0-beta): /api/stream SSE · Three.js fleet mesh · chase/orbit/top · telemetry · Stop/Reset.

python3 gui/server.py --port 8095
cd gui/v3 && npm run build   # or npm run dev
python3 tools/gui_v3_eval_bundle.py   # headless eval → docs/evidence/review/
# http://localhost:8095/v3/#/scenario

0. 2026-06-21 GUI v3 planning — DONE (superseded by scaffold above)

Direction locked: Scenario-first · Run-second · Svelte 5 + Vite (GUI_V3_ADR.md).

0. 2026-06-20 thesis freeze + core → main port — DONE

Tag: vdsim-thesis-beta @ 6623f55 (VLA plant beta before core commits).

Commits on thesis (not all on main): - c4b5c0a — core (#3–9, per-wheel tire) → cherry-picked to main (38b1955 + 5cdca4e) - e5fb2c6 — VDSimPlant wiring → thesis-only

Develop on main going forward; thesis = VLA delivery / hotfix only.

0. 2026-06-18 pre-main cleanup (agreed items 3–5,7–9) — DONE

Shipped (⅙/2 deferred — interface discussion pending): - #4 tire(wheel) — IVehicleDynamics::tire(int) on L1/L2/L5; pybind dyn.tire(wheel=0) - #5 steer lag — vp.steer_deadtime_s → ActuatorModel (direct path uses actuator, removed direct_steer_lag_) - #8 ladder lowering — ladder_lowering.hpp centralises 1500×5 / 600 / 4000 (ISO frozen) - #7 friction canonical — create_friction_patch_ground → wide-y polygon + PolygonFrictionGround; make_friction_ground() + FrictionMapConfig - #3 session factory — make_direct_control_session(vp, TireSetup, sp, opts); make_vla_plant_session = shim - #9 thin plant — SimOutput wheel GT fields; vdsim_plant._obs_from_output(sess.output())

Verify: cmake --build build -j && cd build && ctest --output-on-failure && python3 tests/scripts/test_vla_plant.py

0. 2026-06-18 per-wheel tire (advanced) — DONE

Shipped: FL/FR/RL/RR can each use different TireParams on L1/L2/L3/L5 force loops. Unified (one TireParams) and per-axle (front, rear) constructors remain backward compatible. Catalog optional slots: tire_rear (RL+RR when no corner slot), tire_fr / tire_rl / tire_rr → resolved tire_*.yaml in fleet row.

Deliverables: - TireSetup (params.hpp) — wheel[4] + for_wheel / for_axle; corner ctor + from_corner_yaml_paths - Dynamics: 4× ITireModel via init_wheel_tire_models (was per-axle) - Resolver / materialize: corner yaml paths on ResolvedVehicle + fleet row keys - Cosim: VehicleSpawn.tire_{fr,rl,rr}_yaml + load_tire_setup - pybind: wheel, front/rear properties, from_corner_yaml_paths - Tests: PerWheelTire.L2FrontLeftRightDifferentGrip (+ existing PerAxleTire)

Verify: cmake --build build -j && cd build && ctest -R 'PerAxleTire|PerWheelTire' --output-on-failure

0. 2026-06-18 per-axle tire (core) — DONE

Shipped: front/rear axle can use different TireParams (MF96/MF2002/LuGre/belt) on L1/L2/L3/L5 force loops. Single TireParams still duplicates to all corners (backward compatible). Catalog optional tire_rear slot → tire_rear.yaml in resolved fleet.

Deliverables: - TireSetup — per-axle ctor / from_yaml_paths; shim initialize(vp, TireParams, sp) - Dynamics: per-wheel ITireModel routing (same axle shares params unless corner slot set) - Resolver: tire_rear optional blueprint slot; ResolvedVehicle.tire_rear_yaml - Cosim: VehicleSpawn.tire_rear_yaml → TireSetup::from_yaml_paths - pybind: TireSetup, initialize_setup - Tests: PerAxleTire.L2FrontRearDifferentGrip, PerAxleTire.L1UnifiedMatchesSingleTireParam

Verify: ctest -R PerAxleTire --output-on-failure

0. 2026-06-18 VLA plant P2 — smooth 1-D friction blend + steer lag docs — DONE

Shipped (customer P2, docs/PLANT_CUSTOMER_FEEDBACK.md): - 1-D friction_map smooth transition — FrictionPatchGround linear blend over 1.0 m before x0 and after x1; outside x < x0−1 / x > x1+1 → base_mu; overlapping patches → min μ (same rule as 2-D polygon). - VehicleParams.steer_deadtime_s — 1^st-order steer lag (τ) on VLA direct-control path (SimSession::tick when direct_control_path); default 0 = instantaneous ZOH δ. - Docs: docs/VDSIM_PLANT_TUTORIAL.md §6a (1-D blend), §7 (steer lag manual workflow).

Deliverables: - core/src/contact_providers.cpp — FrictionPatchGround blend - core/src/sim_session.cpp — direct-path steer lag from vp.steer_deadtime_s - Test: VlaPlant.FrictionPatchSmoothTransition

Verify: cmake --build build -j && cd build && ctest -R VlaPlant --output-on-failure

0. 2026-06-18 VLA plant P1 — roll/pitch obs + 2D polygon friction map — DONE

Shipped (customer P1, docs/PLANT_CUSTOMER_FEEDBACK.md): - obs["roll"], obs["pitch"] [rad] — quasi-static angles from roll_angle_qs() / pitch_angle_qs() (L2–L5; L1→0). Already on IVehicleDynamics + pybind; wired in python/vdsim_plant.py _obs(). - friction_map_2d — [{"polygon": [(x,y),...], "mu": 0.5}, ...] on VDSimPlant (mutually exclusive with 1-D friction_map). C++ PolygonFrictionGround + create_polygon_friction_ground() in contact_providers.cpp: per-wheel world (x,y), point-to-polygon boundary distance, linear blend to base_mu over 1.0 m global; overlapping patches → min μ.

Deliverables: - core/include/vdsim/interfaces.hpp — PolygonMuPatch, factory decl - core/src/contact_providers.cpp — PolygonFrictionGround - python/bindings.cpp — PolygonMuPatch, create_polygon_friction_ground, make_vla_plant_session(..., poly_patches, blend_distance) - python/vdsim_plant.py — friction_map_2d, roll/pitch in obs - Tests: VlaPlant.PolygonFrictionInsideAndOutside, PolygonFrictionBlendZone; Python test_roll_pitch_in_obs, test_polygon_friction_map_2d - Docs: docs/VDSIM_PLANT_TUTORIAL.md §6b (2D polygon), §4 roll/pitch

Verify: cmake --build build -j && cd build && ctest -R VlaPlant --output-on-failure && python3 tests/scripts/test_vla_plant.py

0. 2026-06-18 Dynamics ladder bench L0–L5 — DONE

Shipped: tools/bench_dynamics_levels.py + examples/bench_dynamics_levels.cpp (vdsim_bench_levels) — 당시 차량 프리셋(0.7.0에서 제거), dt=5e-4, 500 step, SimSession.tick() ms/step.

level	ms/step (AILAB-12, Release)
L0 kinematic	0.0005
L1 bicycle	0.0162
L2 seven_dof	0.0313
L3 fourteen_dof	0.0317
L4 fourteen_dof_kinematic	0.0321
L5 Free3D stunt	0.0354

L0–L4 via pybind make_sim_session; L5 via C++ binary (not in pybind session factory).

Run: cmake --build build -j && python3 tools/bench_dynamics_levels.py

0. 2026-06-18 VLA plant per-wheel peak slip (alpha_peak, kappa_peak) — DONE

Shipped: per-wheel MF pure-slip peak positions (alpha_peak [rad], kappa_peak [-]) at current Fz/load — raw GT for rising/peak/sliding (drift) classification vs current slip.

Deliverables: - mf_peak_slip(B,C,E) in interfaces.hpp (Newton on MF peak equation) - ITireModel::Output/Wrench::alpha_peak/kappa_peak + IVehicleDynamics::wheel_alpha_peak/kappa_peak() (Ld2) - MF2002 / MF96 / linear compute() fill; evaluate() + Ld2 step propagate - python/bindings.cpp + python/vdsim_plant.py wheel[i]["alpha_peak"], ["kappa_peak"] - Test VlaPlant.PeakSlipMatchesNumericalArgmax

Verify: cmake --build build -j && cd build && ctest -R VlaPlant --output-on-failure

0a. 2026-06-18 VLA plant realized peak mu (useGT option A) — DONE

Shipped: per-wheel mu_peak (load-dependent MF realized peak coefficient, force/Fz) exposed alongside existing contact wheel.mu. Friction-circle GT metric useGT = ||[Fx,Fy]|| / (mu_peak*Fz) is now <= 1 by construction; contact mu preserved.

Deliverables: - ITireModel::Output/Wrench::mu_peak + IVehicleDynamics::wheel_mu_peak() (Ld2) - MF2002 / MF96 / linear compute() fill mu_peak; evaluate() propagates to wrench - python/bindings.cpp + python/vdsim_plant.py wheel[i]["mu_peak"] - Test VlaPlant.RealizedPeakMuBoundsFrictionCircle

Verify: cmake --build build -j && cd build && ctest -R VlaPlant --output-on-failure

0b. 2026-06-18 VLA plant tyre → MF2002 (T-IVb) — DONE

Shipped: the VLA plant preset now uses MF2002 (load-dependent PKY* cornering stiffness + PDY2 peak-μ load sensitivity) instead of MF96.

Deliverables: - the plant vehicle preset, its MF2002 tyre YAML + .tir and an MF96 tyre YAML (all four removed from the public tree in 0.7.0 and kept privately; the MF2002 tests now run on external/chrono_parity/sample_pac02.tir) - python/vdsim_plant.py — resolves relative tir_path from configs root - python/bindings.cpp — exposes TireParams.backend / tir_path to Python - Tests (test_vla_plant.cpp): DryHandlingMatchesLinearBicycleRealMf2002Tire, Mf2002LoadSensitivityConcaveStiffnessAndMu, Mf2002FrictionPatchHalvesPeakGrip, CombinedBrakeTurnGripLossOnPatchMf2002 (+ legacy MF96 inline tests unchanged)

Verify: cmake --build build -j && cd build && ctest -R VlaPlant --output-on-failure

Spec: docs/design/VLA_THESIS_PLANT.md

0c. 2026-06-18 VLA thesis plant (T-IV) — DONE

Shipped: vdsim_plant.VDSimPlant — first external client plant API for closed-loop MPC.

Deliverables: - #0 Ld2 native CmdL1 torque path (Fx→per-wheel τ, bypass throttle map) + plant_path kinematic-blend bypass - #1 create_friction_patch_ground(base_mu, [(x0,x1,mu)...]) + pybind - #2 wheel_mu() on Ld2 + pybind - #3 plant vehicle preset + MF96 tyre YAML (removed in 0.7.0) - #4 python/vdsim_plant.py on vdsim_lab / make_vla_plant_session (direct CmdL1, ZOH two-tier dt) - #5 smokes: tests/scripts/test_vla_plant.py + examples/vla_plant_demo.py + python/VDSIM_PLANT_README.md

Verify: cmake --build build -j && cd build && ctest -R vla_plant --output-on-failure (headline #2 brake+turn on μ=0.5 patch).

Spec: docs/design/VLA_THESIS_PLANT.md

0d. 2026-06-14 세션 인수인계 (archive)

목표였던 것: 멀티차량 비교(헤드리스+GUI) + 다나와식 차량조립 UI + 외부 AI 디자인 리뷰 반영. 완료.

현재 상태: main local (ahead of origin) · 385/385 green · thesis core port done · VLA on VDSim-Thesis only.

이번 세션 main 에 들어간 것: - 헤드리스 멀티차량 비교 tools/vdsim_compare.py — preset N대에 같은 ISO maneuver(step-steer/skidpad/DLC) 돌려 지표 나란히 표 + bar-chart(compare.csv/png). examples/maneuvers.py 함수에 optional veh=(vp,tp) + trace=True(step_steer yaw-rate 시계열). - 다나와식 build-sheet (assembly 모달 재설계) — 좌측 슬롯 리스트(Body&chassis/Powertrain/Grip/Control 그룹, 각 행=부품+change) → 클릭 시 우측 그 슬롯 라이브러리 + RESOLVED STATS. assembly.py 카테고리 새 taxonomy 로 재그룹. - GUI Compare 대시보드 — VEHICLES 패널 ⊟ → blueprint 체크박스 → /api/compare(routes.py, vdsim_compare 래핑) → step-steer yaw-rate overlay SVG 차트 + Δ%(baseline 대비) 표(유효숫자 통일, 정성/정량 분리). tests/scripts/test_compare.py(ctest multi_vehicle_compare). - GUI 폴리싱(디자인 리뷰 반영) — RESOLVED STATS 단위, active 행 회색+좌측 accent, 3D 프리뷰 우측, Sync-draft 표준 크기.

다음 할 일 후보 (우선순위 순): see ROADMAP.md §15.

1. P0 — tag v0.6.1 / v0.6.2 done (v0.6.1 @ 5cdca4e, v0.6.2 @ 9a40c1b).
2. P1 — #1 Fx→τ C++ contract, #6 drive_split / brake bias design.
3. P2 — Chrono KC gen_kc_reference API fill-in done — CSV committed, ChronoKcParity active.
4. P3 — v0.5.2 GUI: terrain Play or .tir import; GUI v2 framework decision.
5. Thesis — VLA hotfix only on VDSim-Thesis; plant → main PR after client sign-off. 다음 할 일 후보 (우선순위 순):
6. GUI 재아키텍처 (전략 결정 후) — 외부 AI 리뷰 4/10 "프로토타입" 판정. 큰 항목: dockable/resizable 패널 + 모달 제거(분할화면), 글로벌 모드 nav(Setup-Sim-Analyze), 전사 디자인 시스템. stdlib http.server + 손코딩 app.js 의 한계 — React/Svelte + 디자인 패스(Figma) 결정이 선행. 지금 찔끔 하지 말 것.
7. #158 GUI 멀티차량 3D render — preset picker 로 spawn 은 됨, 3D 뷰 N대 동시 표시(VDS1 vehicle_id demux) 남음.
8. cosmetic — part 라벨 "Fsk Formula chassis body" 잔재, Session tune 탭 라벨.

주의·함정: - retax2 breaking — 옛 chassis.*/susp.* part id, front_susp_kin 슬롯 제거됨. - from_yaml 은 envelope(body:)를 unwrap 안 함 — IsoBaseline 은 body/sedan.yaml(→struct defaults) 로딩. 건드리지 말 것. - ISO/Default subsystem 숫자 리베이스 금지. 영상 git 커밋 금지. 현대 .tir 대외비. - GUI 서버 띄울 때 run_in_background 사용. pkill -f "gui/server.py" 는 자기 명령줄 self-match 로 죽으니 pkill -f "server.py --port" 처럼 좁혀서. - 멀티차량 = 비교용, V2V 충돌 scope 밖 (메모리에 기록됨).

관련 경로: tools/vdsim_compare.py·examples/maneuvers.py (비교) / gui/static/app.js(openCompareModal/renderCompareChart/renderAssemblyPane)·gui/api/routes.py(/api/compare)·python/catalog/assembly.py (GUI) / cosim/realtime_server.cpp (멀티차량 런타임).

1. 문서 인덱스

Topic	User / ops	Design spec
Catalog + scenes	CATALOG_AND_PHYSICS.md	design/PARTS_CATALOG.md
Ld4 multibody M1–M7	§ L4 in catalog doc	design/LD4_MULTIBODY.md
v0.4 slope / jump / loop	RUNNING.md	design/V0.4_SLOPE_JUMP_DYNAMICS.md
v0.5 terrain + L5	—	design/V0.5_TERRAIN_L5.md
LuGre tire	theory ch.19	design/V0.2_TIRE_LUGRE.md
Roadmap	ROADMAP.md	design/TIRE_ROADMAP.md
Validation	VALIDATION.md	—

2. 브랜치 스냅샷

v0.4.0 stunt (Ld5) — shipped + tagged

• create_stunt_dof() → Free3DDynamics (free_3d_dynamics.cpp): pos+quat+ω, 4-wheel MF, no rail snap. 유일한 stunt plant — create_legacy_stunt_dof()(L3+L2 rail loop)는 cleanup 에서 제거됨.
• Scenes jump_ramp_demo.yaml / vertical_loop_demo.yaml; cosim stunt: → make_ground; GUI stunt mesh + z/pitch/roll telemetry; settle_spawn_on_ground.
• LuGre default retune for L5 loop (sigma0 9e4, sigma2 75). Theory ch.20.

v0.5.0 terrain + L5 — shipped + tagged (headless/batch/cosim)

• Hub contact 통일: Flat/SplitMu/Inclined/Rough/Heightmap → wheel_world_positions + hub_penetration.
• CurvedGround banked turn (create_curved_ground, cosim stunt.ground == banked).
• Scenes terrain_hill_demo / banked_grade_demo / banked_oval; tools/bake_synthetic_hill.py + assets/terrain/hill_demo.bin; materialize terrain: forwarding.
• Tests tests/integration/test_terrain_l5.cpp (7) + test_l5_driving.cpp.
• v0.5.2 deferred (browser): M4 GUI terrain Play, M5c GUI stunt 저작.

GUI cleanup (이번 세션)

• app.html 4873→~250L shell; inline CSS/JS → gui/static/{app.css, app.js, util.js, minimap.js, fields.js, manual.js}; /static/ route.
• legacy scene loaders 제거, low-speed 상수 → vdsim/low_speed.hpp. 검증: esbuild bundle 게이트(node20 @ ~/.nvm).

Ld4 multibody (M1–M7) — shipped

상세: design/LD4_MULTIBODY.md

M	Runtime	Summary
M1	L4	FK attach, create_fourteen_dof_kinematic()
M2	—	Topology graph from kin YAML
M3	design	Quasi-static bushing compliance
M4	L4 step	Hard-joint DAE (TA/MP/DW/5-link) + Baumgarte travel
M5	GUI/py	K&C sweep charts
M6	CLI/py	Adams CSV import x-check (5 gains, 5% rtol)
M7	offline	Lumped 1-DOF revolute RNEA (TA/MP/DW/5-link)

GUI / cosim (v0.2 carry-over)

• Vehicle Edit: Chassis / Tire / Suspension tabs; FL FR RL RR
• VDS1 protocol v5 (throttle_applied, brake_applied)
• L4 in make_dyn("L4"), catalog, l4_sedan_kinematics.yaml

3. 빌드·검증

cmake --build build -j
cd build && ctest --output-on-failure   # 388/388 (main)
python3 gui/server.py                   # http://127.0.0.1:8080

4. 다음 작업

Current sprint: ROADMAP.md §15 (development priorities).

Shipped on main (2026-06-20): SessionKind, KC compliance, T3 VLOW. 388/388 ctest.

Deferred (v0.5.2, browser): GUI terrain Play, stunt authoring, .tir import; tire-force arrows in wheel frame in 3D view.

Historical archive (v0.4–v0.5 tire stack)

[`design/TIRE_ROADMAP.md`](design/TIRE_ROADMAP.md) §0/§4 + theory ch.21. - **T1 MF2002**: `TireParams.backend` ("mf96" default | "magic_formula" | "linear") + `tir_path`; `create_tire_from_params` dispatch in every dynamics `initialize()`; MF2002 combined slip bypasses the host friction ellipse. `.tir` stays uncommitted (gitignore + confidential). Tests `Mf2002Catalog.*`. - **T2 belt transient**: `vdsim/belt_tire.hpp` `belt_relax()` (tau=sigma/|Vx|, exact exp); opt-in `TireParams.belt {enabled, sigma_lat, sigma_long}`; wired in seven_dof (L2/L3) + free_3d (L5), both MF (relax kappa/alpha) and LuGre (relax slip velocity). Tests `BeltTire.*`, `BeltTransient.*`, `BeltValidation.*`. **Default off -> no drift.** - **Decision (locked):** keep own lean tire stack; Chrono Pac02 (BSD-3) is a cross-validation *reference*, not a dependency. Belt = own (no permissive OSS exists). - **Chrono Pac02 parity (DONE):** isolated cross-check in `external/chrono_parity/` (Chrono 8.0.0 built from source at `~/build_ext/chrono_src`; never in VDSim's CMake). Generator `gen_reference.cpp` (ChTireTestRig) -> committed `reference/pac02_reference.csv`; VDSim gate `ctest -R ChronoPac02Parity`. **Pure-long Fx ~2% AND pure-lat Fy ~1%** vs an independent Pac02 — both gated (backbone + load sensitivity + cornering stiffness validated). Combined cross-terms differ but that is largely a rig-frame artifact (`ChTireTestRig` reports global-frame force on a yawed wheel) mixed with weighting — report-only. See `external/chrono_parity/README.md`. - **Tire T2 complete:** bicycle (L1) belt wiring done (MF + LuGre) — belt now on L1/L2/L3/L5, default off. Tests `BeltTransient.L1*`. **Remaining tire:** GUI `.tir` import (v0.5.2 GUI). (Combined-slip Pac02 alignment is low priority — pure axes already match; the residual is mostly the rig-frame decomposition, not our model.) **v0.5.2 (deferred — needs browser, no headless path):** 1. **M4 GUI terrain load + L5 Play** — chase cam uses `position.z`, spawn on mesh. 2. **M5c GUI stunt authoring** — author ramp/loop/banked scenes (today render-only). 3. GUI `.tir` import (Tire T1 tail). 4. **Tire-force arrows in wheel frame** — model now exposes `tire_forces_wheel()` (un-rotated, validated by `TireFrame.*`; also in pybind). The GUI still draws the body-frame `tire_forces_body()` along body axes (arrows don't tilt with steer); to show tire-frame, send the wheel-frame force over VDS1 and parent the arrows to the steered wheel mesh (`gui/static/app.js makeForceArrow` / update L793-794). **Done this session (post-tire):** - ISO re-baseline (flat, sedan L2 LuGre) — VALIDATION.md table refreshed + CI gate `tests/integration/test_iso_baseline.cpp` (`ctest -R IsoBaseline`) locks the 7401 signature. - **Drivetrain v2 (opt-in `powertrain:` block, L2/L3):** 2D engine torque map + N-speed gearbox (RPM coupling, gear-dependent reflected inertia, idle-floor/launch clutch) + shift policy (manual / auto-RPM / **user-defined function**, pybind callback). Accessors `engine_rpm()`/`current_gear()`. `powertrain.hpp` `EngineGearbox`, `EngineGearboxDrivetrain` in default_subsystems. Theory ch.22, sample `configs/powertrain_sedan_demo.yaml`. Default flat torque unchanged -> ISO green. Tests `EngineMap.*`/`PowertrainYaml.*`/`EngineGearbox.*`/`DrivetrainV2.*`. **317 ctest.** **Remaining:** catalog `drivetrain_v2` part (materialize powertrain into the chassis); L1 + engine-map GUI workshop. **User-defined subsystem modules (this session):** any built-in subsystem can be replaced by a C++ or Python subclass — `BrakeModule`/`SteeringModule`/`DrivetrainModule` (L2/L3/L4/L5), `SuspensionModule`/`AntiRollBarModule` (L3/L4). Install via `model.set_*_module()` (`IVehicleDynamics`, mirrors `set_shift_policy`; seven_dof + free_3d hold brake/steer/drivetrain, fourteen_dof holds suspension/ARB + delegates the rest to `inner_`, L4 inherits from it). L1 computes brake/drive inline -> hosts none. Subsystem members are `shared_ptr` so a pybind-trampoline Python subclass stays alive. Cadence: `begin_step` once per step, `apply/wheel_torque/force` per RK4 stage; brake/drivetrain torque is **signed** (opposes spin). Tests `UserModules.*` (transparency L2/L3 to 1e-9 + effect + scoping), sample `examples/user_brake_module.py`, theory **ch.23**. **321 ctest.** **Remaining:** GUI module workshop; L1 has no subsystem objects (out of scope). **v0.6+:** 1. Ld4 v0.6 — shared inertia helpers **(done)**; full loop dynamics / Featherstone in step deemed not needed (hard DAE already solves loop kinematics — see LD4_MULTIBODY.md §7). 2. Drivetrain: catalog `drivetrain_v2` part **(done)** + L1 powertrain **(done)**; engine-map workshop (UI) remaining. 3. Tire: T2 belt is **complete** (L1–L5); higher belt eigenmodes (rigid-ring/FTire) out of scope.

5. 주의

• Hyundai (TUR) .tir confidential — 커밋 금지
• Default subsystem 숫자 리베이스 금지
• Wheel: FL=0, FR=1, RL=2, RR=3 · ISO 8855 RH
• Cosim plant v5 ↔ rebuilt vdsim_realtime
• configs/.resolve_cache/ — 커밋하지 않음

6. Git

• main local, ahead of origin · 388/388 ctest · tags v0.6.0 (MBD) · v0.6.1 (core port) + v0.6.2 (quality/T3) after commit/tag.
• External build (not in repo): Chrono 8.0.0 source at ~/build_ext/chrono_src (vehicle module only) — only needed to regenerate the parity CSV, not to run the gate.
• Next: new feature branch off main.

7. Next pickups (fresh session)

See ROADMAP.md §15. Summary: tag v0.6 → Fx→τ contract → KC compliance → GUI v0.5.2 pick-one.