GitHub Copilot Skill

dayz-vehicles

Author, import, debug and ship DayZ ground/water vehicles: CarScript cars, trucks, quads/ATVs and motorbikes plus Boat-based watercraft. Covers the config.cpp/model.cfg contract, drivetrain/engine/suspension, Axles/Wheels, Crew/get-in/actions, AnimationSources, structural parity,

LLM Mart · 0 points · 2 views 0 listing impressions 0 install-command copies
Download willy92wins-dayz-modding-knowledge-pack-skills_dayz-vehicles-22c35fd.zip · 359 KB

Install

skills CLI npx skills add https://github.com/willy92wins/DayZ-Modding-Knowledge-Pack/tree/main/skills/dayz-vehicles
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install willy92wins-dayz-modding-knowledge-pack@llmmart
Git git clone https://github.com/willy92wins/DayZ-Modding-Knowledge-Pack.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole willy92wins/dayz-modding-knowledge-pack collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

DayZ Ground Vehicle Modding

Family selector — day-0 critical path (CAMBIO-2)

  1. Identify family by asset origin and architecture; not by symptom.
  2. If asset is a new Grub source-game car, proxy-split and with moving parts, use family B.
  3. Family B → open ../rip-vehicle-import/SKILL.md and follow only its adapter, golden, and allowlist.
  4. That adapter opens asset asset-contract.json as third and final day-0 file; schema, Blender export, and primitive are machine inputs, not additional documents maintained by agent.
  5. If a frozen plan/runbook already exists for asset, follow that contract in flight; do not migrate it here.
  6. If no row applies or family adapter is missing: STOP. Do not improvise from this atlas.
Input signal Adapter Day-0 files
source-game Grub, new car, proxy-split + doors/moving parts Family B This router → ../rip-vehicle-import/SKILL.md → <asset>\asset-contract.json
Asset already in flight Frozen project runbook Whichever cited in its current plan
Other family or insufficient data None STOP

Rest of this body is diagnostic atlas and does not form part of day-0 critical path. Only consulted when adapter or symptom links a specific section.

Preflight before re-importing / re-assembling any piece (SP-193): first measure FACTORY FACE-LEVEL completeness of piece against already packed model. A "rotten" mesh is usually complete and defect lies elsewhere; see SP-193 at end of this atlas.


Drivable wheeled vehicles in DayZ extend CarScript — there is no generic "vehicle" base, and you do not build one from EntityAI or Inventory_Base. CarScript gives wheels, seats, sound integration and the damage system for free; you configure the rest. This skill is the vehicle-specific layer on top of dayz-model-pipeline (generic geometry / LODs / textures) and is the ground counterpart to dayz-aviation (anything that flies belongs there).

Scope: cars, trucks, quads/ATVs, motorbikes — anything wheeled and drivable, modeled from scratch or imported (Blender / OBJ) from another game.

REDIRECT CAMBIO-1: header family selector and chosen adapter are the sole day-0 path.

Vehicle type matrix (invariant): Car and Boat are SIBLINGS, both directly under Transport (car.c:87 / boat.c:28) — NOT parent/child. Transport owns crew/get-in/flip/fuel; Car owns wheels/brakes/CarFluid; Boat owns propeller/buoyancy/BoatFluid(fuel-only). A truck is a plain CarScript config with 3 axles + double wheels, NOT a new class. Boats, truck double-wheels, ATV and the motorbike gap → references/vehicle-types-boat-truck.md (historical: since DayZ 1.30 Exp motorbikes are a vanilla family, see the dayz-motorbikes skill).

DayZ 1.30 Exp (build 1.30.164014) changes the matrix (added 2026-09-16, source_verified on the extraction at E:\DayZ-Exp-Extract\1.30.164014\exp):

  • [EXACT] A third sibling exists: class Motorbike: Transport with simulation = "motorbike" (dta\bin\config.cpp:1054-1059; engine proto 3_Game\Vehicles\Motorbike.c:30), scripted by MotorbikeScript (4_World\Entities\Vehicles\MotorbikeScript.c:53), which does NOT inherit CarScript. Two-wheel work goes to the dayz-motorbikes skill; this atlas stays for cars, trucks, quads and boats, plus the ViewPilot and the first-person view of any family (SP-248, SP-249 and references/proxies-and-get-in.md).
  • [EXACT] CarScript was refactored around components owned by Transport: VehicleLightsComponent (CarScript.c:187, 229-233), VehicleHornComponent and VehicleVFXComponent (CarScript.c:189, 343-368). The old overrides are marked [Obsolete]: CreateRearLight/CreateFrontLight (CarScript.c:2937-2945), IsVitalCarBattery/IsVitalTruckBattery (CarScript.c:2811-2818, replaced by Transport.NeedElectricitySourceDevice() at Transport.c:830), CarPartsHealthCheck (CarScript.c:2976-2979, replaced by Transport.PartsHealthCheck() at Transport.c:440 / override at CarScript.c:2121). The horn actions are ActionVehicleHornShort/Long (CarScript.c:2319-2320). Contact data is VehicleContactData.m_Impulse (Transport.c:1153-1157, used at CarScript.c:1291-1292). Wheel classes gained tyreRoughness, tyreLongitudinalFriction, tyreLateralFriction (DZ\vehicles\wheeled\config.cpp:749-751; absent in 1.29).
  • [DESIGN] A 1.29 car mod that overrides any of the obsolete members compiles against the attribute but its override is no longer called; re-verify each override against the 1.30 CarScript.c before shipping. Migration checklist and the "1.29 statements that are now false" table: dayz-motorbikes/references/vehicle-1.30-refactor.md. Sections of this atlas that predate 1.30 (lights via CarLightBase, CarPartsHealthCheck, data[0].impulse) describe 1.29 and are not yet rewritten.
  • [EXACT] Transport.GetWeightCoef() default 1.0 (Transport.c:478-481); vanilla Truck_01_Base 2.0 (Truck_01_Base.c:120-123), Offroad_02 1.5 (Offroad_02.c:133-136), Motorbike_01 0.5 (Motorbike_01.c:84-87), Motorbike_02 0.75 (Motorbike_02.c:74-77). Digest M's claim that MotorbikeScript.c:661 returns 0.25 is false: that line uses GetWeightCoef(), it does not define it.
  • [EXACT] CarScript.HasGangedTailAndBrakeLights() default false (CarScript.c:468-471); true on Offroad_02 (Offroad_02.c:68-71) so reverse/position swaps do not kill the brake selection (CarScript.c:2061-2094).
  • [EXACT] Vehicle light toggle is UAToggleVehicleLights / ToggleVehicleLightsActionInput (ActionInput.c:849-855, ActionSwitchLights.c:16-18); UAToggleHeadlight is headgear (ActionInput.c:776-782). ForceUpdateLightsStart/End are [Obsolete("1.30: no replacement")] (CarScript.c:2805-2809).
  • [EXACT] Get-in calls CanGetIn() then IsAreaAtDoorFreeDiag and does call CanReachSeatFromDoors(..., 1.0) from script (ActionGetInTransport.c:36,54,64).
  • [EXACT][CLAIM-VEH-130-PORT-BLOCKERS] Two 1.30 changes break a 1.29 car outright, beyond its looks. CarScript.ToggleHeadlights() was deleted without an [Obsolete] stub (1.29 CarScript.c:2088), so an override of it stops the 4_World compile. A car that does not override the three electricity-source methods reports NO_BATTERY with a battery attached (1.30.164200 transport.c:835-838, carscript.c:1768-1774): no engine start and no light switch. Two more showed up in game on 1.30.164200: a backslash script module path loaded none of the LFQuad's scripts, silently, and every wheel class without a <class>_Ruined twin raised a script error when its first wheel was created (1.30.164200 inventoryitemtype.c:17-46). Detail and the override check: references/dayz-1-30-vehicles.md (table and the path claim below it, section 2, section 11, checklist items 0 and 11-13). Seen porting the LFQuad to 1.30.164200 (2026-10-26).
  • Lights, horn, fluids, boats, wreck cisterns-as-Well, speedGeoms, and the motorbike 3PP camera pointer: references/dayz-1-30-vehicles.md. Motorbike camera ID 32 lives in dayz-motorbikes/references/rider-animation.md.

Per-tick event asymmetry (invariant, added 2026-07-29, LFHeli OH-1): CarScript has NO EOnSimulate — its ctor registers only POSTSIMULATE/POSTFRAME (carscript.c:207-208) and ALL its per-tick work (engine, fluids, part health, fuel/engine auto-stop, exhaust and wheel FX, much of it gated IsServerOrOwner() — vanilla runs it on the owner client, but in mods replica-dependent checks (CheckVitalItem) must stay server-only (LL-305)) lives in EOnPostSimulate (carscript.c:807). BoatScript overrides BOTH (boatscript.c:355 and :393). Consequences on a car: hook EOnPostSimulate for per-tick logic; super.EOnSimulate is the empty stub at enentity.c:201-203, so calling or skipping it does nothing either way — treat any comment claiming it "runs engine/fluids twice" as false. Custom solvers that add SetEventMask(EntityEvent.SIMULATE) and pump it by hand (aviation) → dayz-aviation preflight invariants. (hasta 1.29: those carscript.c line numbers.) (desde 1.30: SetEventMask is CarScript.c:207-208; EOnPostSimulate is CarScript.c:807. Full line-drift table: references/dayz-1-30-vehicles.md.)

Cockpit screen → live engine map (RTTextureWidget bridge) (invariant, added 2026-08-19, SUB_BRZ RT spike): Any hiddenSelections screen can display the ENGINE's map renderer (what LBmaster's GPS uses) instead of baked tiles: create the RT via CreateWidget(RTTextureWidgetTypeID, …) (children render INTO its texture, enwidgets.c:28 — they never paint to the viewport, so no HUD witness), hang a MapWidget under it with CreateWidgets(layout, rt), then hook the entity in this exact order: SetObjectMaterial(idx, "Mod\data\x.rvmat") FIRST (no leading backslash), SetObjectTexture(idx, "$rendertarget") second, proto native void SetGUIWidget(IEntity, int idx, RTTextureWidget) (enwidgets.c:637) LAST. A later SetObjectMaterial rebinds the base texture slot and silently undoes the hook — a uniform light-grey screen means $rendertarget resolved to the white fallback texture. Zero call-sites exist on disk for SetGUIWidget/RTTextureWidget; verified in-game on 1.29 diag. MapWidget GetScale() defaults to 0.85; at 0.05–0.6 it draws line-work only (contours + roads, no urban/forest fill) — pick the zoom with the user's eye gate. If the hooked vehicle leaves the client streaming bubble its GetPosition() reads <0,0,0> and the map recenters to the map corner: production code must re-resolve the vehicle every tick, never cache the entity reference. MCP test flow that actually seats and starts: vehicle_get_in_client (ownership get-in; vehicle_enter answers seated:1 without sticking) → engine_set. Recipe, captures and gotchas: LFVehicleUI_dev\HANDOFF.md snapshot 2026-08-19. Closed 2026-08-19b — RT unusable for a live GPS: the RT freezes the FIRST composed frame (later SetScale/SetMapPos/vehicle movement never recompose it), and a MapWidget child only ever contributes the vector layers (roads + contours) — the raster fill (land, forest, buildings) never reaches the RT, even with a 24 s warm-up between widget creation and the entity hook. The bridge recipe above remains valid for STATIC screen content only. Live vehicle GPS goes through baked tiles on hiddenSelections instead (next invariant).

Vehicle screen tiles: terrain satmap, not navigation usermap — and apply server-side (invariant, added 2026-08-19, SUB_BRZ Plan B): Two tile sources share the same s_XXX_YYY_lco.paa naming and the same 32×32 grid of 480 m tiles on chernarusplus, but look completely different: dz\gear\navigation\data\usermap\ is the PAPER map art (light background, red roads, contour lines — what the ItemMap shows), while dz\worlds\chernarusplus\data\layers\ is the colored aerial satmap (the real "GPS look"). A GPS screen wants the layers path; world name must be parameterized per terrain. Second half of the invariant: SetObjectTexture/SetObjectMaterial called ONLY client-side is silently reverted to the config hiddenSelectionsTextures[] whenever the server pushes a visual-state resync — observed reproducibly on GEAR CHANGE. Applying on the dedicated server does NOT fix it (refuted in-game 2026-08-19 afternoon): scripted SetObject state does not replicate server→client.* The working pattern is the skin-mod one: re-assert the full scripted visual state client-side on EVERY OnVariablesSynchronized (no packed==applied early-out; gate only the diagnostics). Also force-hide every bank window on the first apply (SetAnimationPhaseNow(name, 1.0) loop): windows spawn with the config-default texture and their initial hide phase is not guaranteed applied. The force-hide only exists at runtime if the hide animations actually BAKED — see the skeletonBones invariant below. Satmap-in-bank verified in-game 2026-08-19 (SUB_BRZ, 1.29 diag, human driver gate).

Tile overlap: neighboring satmap layer tiles share 32 px — sample only the tile core (added 2026-09-14, SUB_BRZ s93). [EXACT] In dz\worlds\chernarusplus\data\layers\, column 480 of a tile is column 0 of its east neighbor and row 480 is row 0 of its south neighbor: a per-tile screen map must sample only the core 16/512..496/512 of each tile. Without the crop the measured seam ratio was ~1.9-2.1 and roads jumped ~30 m; with it, 0.90-1.12 (measured offline on the 1.30.164014 Exp satmap tiles). [DESIGN] Name index of s_AAA_BBB_lco.paa: AAA is the tile column tx, BBB encodes the row as 31 - tz, and UV (0,0) is the northwest corner.

Sliding-plate tile window: fix the border row and size the bezel with wy + COVER + T (SP-402, added 2026-10-01, SUB_BRZ s95). [DESIGN] On a sliding-plate tile map with the marker centred, the hidden margin under the window must reach one full tile past the lowest window corner: with a free bottom row it is 1.5 tiles deep (97.5 mm under the centre on a 7x5 grid of 32 mm tiles, window +/-88.63 x +/-52.92 mm, bezel 99 mm). Pinning the bottom row with minValue/offset0 in its _y class (same slope), whenever its top edge stays COVER below the lowest projected window corner, cuts the requirement to 88.68 mm and the bezel to 90 mm. Under the window the requirement is always wy + COVER + T; with a centred marker nothing below one full tile reduces it. If a smaller bezel is wanted, the lever is tile size (zoom), not the window - say so with a preview.

Seated-view geometry lives in the View Pilot LOD (invariant, added 2026-08-19, SUB_BRZ v1.2): A car with a View Pilot LOD (resolution 1100) renders THAT copy of the geometry for seated occupants, not LOD0. Any geometry/UV/selection patch that must be visible from the driver seat (screen UV remaps, marker quads, interior fixes) has to be applied to BOTH res 0 and res 1100 — patching only LOD0 looks fine from outside and unchanged from the seat, which reads as "the fix did nothing". Seat-only geometry (1PP gap closures) goes in res 1100 alone (SP-469). Also from the same session: a tiny 8×8 DXT PAA is invalid (mip chain below 4 px → engine renders the grey fallback); for flat-color config slots use a procedural #(argb,8,8,3)color(r,g,b,1,CO) instead of a file.

Animations only bake for selections that are skeletonBones (invariant, added 2026-08-19, SUB_BRZ grey-screen root cause): AddonBuilder silently DROPS every model.cfg animation whose selection is not declared in the model's CfgSkeletons skeletonBones[]: the build exits 0, the anim classes stay in the source model.cfg, the AnimationSources stay in config.cpp, and SetAnimationPhaseNow on their sources becomes a silent no-op. Measured differential (SUB_BRZ): gps_marker was a bone → its 4 anims baked into the ODOL; the 28 nav_w* windows were in CfgModels sections[] but NOT bones → all 28 type=hide anims dropped, so the untextured bank could never be hidden and rendered as an opaque grey panel 0.5–1.85 mm in front of the real screen — a grey that reads as "my texture apply fails" while the apply is actually landing fine behind it. Two rules: (1) every animated selection — hide anims included — must be listed in skeletonBones[]; being in sections[] is NOT enough. (2) Verify the bake OFFLINE before deploying: extract the PBO and string-scan the binarized .p3d for the animation SOURCE names, which are plaintext in the ODOL (e.g. re.findall(rb'nav_w\d\d_h', odol) must yield all 28 distinct names). Zero hits = the animation does not exist at runtime, whatever model.cfg says. Debugging corollary: the engine grey fallback is per-SECTION (texture_index=-1); when a screen shows uniform grey, check for coplanar untextured geometry IN FRONT of the selection being textured before blaming the texture path or the apply call.

Verify new geometry by its ODOL SECTION, not by strings (invariant, added 2026-08-19, SUB_BRZ marker hunt): Geometry added to a MLOD by tooling can be present in every string table of the binarized p3d — selection name, material, texture — and still never render, because whether the engine draws a section is decided by fields the strings do not show. Measured on the SUB_BRZ dash marker across seven builds:

Section special Meaning Renders?
0x0002C000 hidden-selection + opaque pass yes (all working textured sections)
0x0002C100 hidden-selection + alpha pass alpha-dependent, unreliable for tiny quads
0x0000C000 no hidden-selection bits never

Two authoring rules follow. (1) A bone-skinned textured selection must ALSO be listed in CfgModels.sections[], or binarize emits is_sectional=0 / sections=[] and there is no draw call for the bone — being a skeletonBones entry is necessary but not sufficient. (2) The face texture must be a real .paa FILE: a procedural #(argb,...) face texture on new geometry produces 0x0000C000. Procedurals are fine in rvmat stages, not as the face texture of a section you need drawn.

Gate before every deploy that adds geometry: extract the PBO and parse the ODOL section table, asserting is_sectional==1, non-empty sections, and the 0x20000 bit. The LOD address table is found by chaining (lod_end[i] == lod_start[i-1], max end ≈ file size); a naive scan for the first plausible table hits a false positive. Reference implementation: <vehicle-import>\work\navscreen_planb\gate_sections.py (built on an external ODOL→MLOD converter's odol_reader). String-only scans pass while the geometry is undrawable — that gate cost five wasted in-game cycles before it existed.

PAA files: always encode with ImageToPAA (added 2026-08-19): A hand-rolled PAA whose mip chain runs below 4 px is invalid DXT. The failure mode depends on the pass: an 8x8 DXT1 _co renders as the engine grey fallback, while a _ca with 2x2/1x1 mips renders as nothing at all on an alpha section. Encode every texture with DayZ Tools\Bin\ImageToPAA\ImageToPAA.exe from a >=64 px source and let it build the mips.

Recycle existing faces instead of authoring new ones (invariant, added 2026-08-21, SUB_BRZ marker): On an imported/ripped vehicle model, faces ADDED by tooling can be perfectly formed and still never render, while every modification of pre-existing faces shows up reliably. Measured over ten builds on SUB_BRZ: a static witness quad — no bone, no animation, bright file texture, section 0x0002C000, winding/flags/normals identical to neighbouring faces that render, and nothing occluding it — stayed invisible in-game; a bank-window quad that already existed was shrunk, retextured and re-UV'd in the same file and rendered immediately. When a model needs a new visible element, look for geometry already in it that can be repurposed (an unused overlay quad, a hidden variant, a duplicate panel): you inherit its section, its bone and its proof of rendering. Reserve authored geometry for models you build from scratch, and if you must add faces to an imported model, prove one renders before building anything on top of it.

Translation offsets: make the memory axis exactly 1 m and the ambiguity disappears (added 2026-08-21): For type="translation" in model.cfg, offset0/offset1 are documented inconsistently: they behave as metres in some sources and as a fraction of the axis vector in others, and picking wrong is silent — the selection simply moves somewhere off the model, which reads exactly like "my geometry does not render". SUB_BRZ lost six build cycles to a marker translated with offset1 = 1.0 against a 169 mm axis: under the metre reading each phase step threw it half a metre out of the cabin. Author the memory axis 1.0 m long in the travel direction and set offset1 to the travel in metres (0.169 for 169 mm of sweep). Both readings then produce the same displacement, so the question never has to be answered. When a hidden/animated selection is invisible, disconnect its animations and rebuild before blaming geometry or materials — a detached control build is one cycle and it splits the two causes cleanly.

Reading a dark-art dash at night (added 2026-08-19): Instrument-cluster art driven by dashboardMatOff (emissive 0) is invisible in an in-game night, which reads exactly like "the texture never applied". Pin server time (serverTime="YYYY/M/D/12/0" in serverDZ.cfg) before judging any interior art gate, and give the OFF material a faint emissive so the furniture stays readable after dark.

Crew / engine state on a non-owner occupant (SP-288, origen LFHeli LF-008): Car.EngineIsOn() on a client that is not the Pawn owner returns false even when the authority has the engine running — no exception, no warning. Measured on the occupant of seat 1 with a two-component probe in the same frame (natEng=false syncEng=true) and the server log showing no stop. Any client subsystem gated on EngineIsOn() shuts itself off (thermal camera died instantly: LF-008). Remedy: replicate the state (RegisterNetSyncVariableBool written in OnEngineStart/Stop) and have every client path read native OR replicated; an m_EngineOnAuthority (OwnerState) snapshot does NOT serve passengers. When a client check decides to turn something off, print BOTH sources in the same frame.

Scripted-pilot limitation — UAInput has no injection API (SP-276b, added 2026-08-17, LFHeli LF-001): the dayz-mcp bridge cannot pilot a vehicle whose mod neutralizes the CarController (SetSteering(0) every tick) and reads its own UA axes, because UAInput does not admit value injection. The working path is a DIAG script in the OWNER that substitutes the UA reading before WriteToMove: with the Pawn pattern of active forces the server consumes the axes from the Move (ConsumeMove → ReadFromMove), so the script needs no sync and no RPC. Plan v0: LFHeli_dev/plans/2026-08-17-piloto-scriptado-diag-v0.md. Status: pending — conditioned on the scripted pilot actually flying (LL-218: do not promote a mechanic without in-game adjudication); SP-276 patch (a) was adjudicated in-game 17-Aug.

Mirror-vs-rotation invariant (added 2026-08-02, LFHeli HH-60G): before you declare an imported model "mirrored" and reach for the det=-1 import rule, check all THREE axes. A table of memory points cannot tell a reflection from a rotation on its own — points have no handedness, so only the PATTERN of sign flips does:

The general test is the SIGNED VOLUME, not counting axes. Take four LABELLED, non-coplanar corresponding points (A,B,C,D) and compute det[B-A, C-A, D-A] on each side. An individual point has no handedness, but a labelled tetrahedron does:

  • signed volume keeps its sign -> proper transform (det=+1) -> rotation. The model is FINE and det=-1 would ADD a mirror that is not there.
  • signed volume flips sign -> improper transform (det=-1) -> the reflection is real; apply the det=-1 import rule plus pseudovector (-S.n) normals.
  • volume ~= 0 -> your four points are coplanar: pick others, the test says nothing.

Counting inverted axes is only the axis-aligned shortcut of that test, and it is easy to misread:

axes flipped (axis-aligned case only) what it is
ONE inverted, other two intact reflection
TWO inverted, third intact rotation of 180 deg about the third axis
all three inverted point reflection = 180 deg rotation + mirror; decompose before acting

Real case, and it nearly cost a full re-emission of 16 .p3d: one session measured only X, saw the sign flipped and concluded "mirrored"; the user authorised de-mirroring. Measuring Z too showed 38/38 points with X and Z inverted and Y intact — a 180 deg yaw (det=+1), which is the pipeline's normal convention, not a defect. Control that settled it: the sister airframe that already flies in-game has the SAME convention (nose at -Z in the file). Applying det=-1 would have introduced a real reflection where none existed and broken a winding policy already verified in-game. The review that settled it did exactly the tetrahedron test: +6.076679 in file against +6.076514 in runtime, ratio +0.999973 - same sign, proper transform, argument closed without a game cycle.

Say what the test does NOT cover. A proper file->runtime transform does not rule out a mirror baked in EARLIER (before the .p3d, in the source mesh or the exporter) nor a pre-mirrored UV island, which is a texture defect and shows up as backwards lettering with the geometry perfectly correct. Those are separate investigations with separate evidence; do not let a clean parity check close them.

Corollary worth more than the rule — prefer RELATIVE measurements. A comparison made within one file, through one reader is robust to whatever axis convention your reader uses; an absolute file-vs-runtime comparison is not. "The 10 crew proxies carry a frame rotated 180 deg from the 15 piece proxies" is sound even if you never establish which handedness py3d reports, because both sides went through the same lens. When a diagnosis can be phrased as a relative comparison against a control inside the same asset, phrase it that way and it survives being wrong about conventions.

Imported-mesh budget invariant (added 2026-07-29, LFHeli HH-60G): a triples/triangle ratio measured on a DECIMATED mesh does not extrapolate to an authored one. Decimate collapses edges and breaks normal sharing, so it inflates resolved (point, normal, uv) triples per triangle. Measured on the same asset: 1.301 on the Decimate-ratio export (0.26-0.6 per group) versus 0.941 on the source's own Medium LODs — 28 % apart, and enough to change how many sub-.p3d you plan for (2.43 projected LODs against 1.75 measured). Rule: measure the ratio on the geometry you will SHIP, not on whatever export you have lying around. Corollary, and it is the bigger win: if the source ships an authored LOD ladder (GTA .yft High/Medium/Low, ripped racing-game, Sollumz-imported .blend datablocks), use it instead of decimating — swapping obj.data to the <piece>_medium datablock is usually better in triples AND satisfies the "the agent never decimates visual" clause of a product spec. Re-measure per model; never copy a ratio or a resolved ceiling between assets.

Packaging invariant — a fix that lives in the generator and not in the binary does not exist (added 2026-07-29, LFHeli HH-60G): when the cycle's fix is adding or changing a PROPERTY of a .p3d (autocenter, class, sbsource, lodnoshadow), a SHA-based manifest cannot see it — it compares the same wrong file on both sides and reports green. Real case: autocenter=0 was added to the hull's visual LOD in both assemblers and the PBO was packed minutes later from .p3d built before; measured afterwards on the 48 extracted models, the 8 sub-models carried it and all 40 hulls did not. The post-ExtractPbo gate must READ the property, per LOD, on the extracted files. Two riders that make the gate survive contact:

  • Declare the expectation in a table with a reason per row, never as "all LODs". A healthy pack contains controls that exist in order NOT to change (an A/B control arm, a byte-frozen baseline); a blanket gate rejects them and gets switched off on its first run.
  • The producer must satisfy its own gate. Here "all LODs" would have rejected the canonical assembler's own output, because the Memory LOD has no geometry to re-centre and therefore never carries autocenter. Exclusions are declared, not assumed.

Deployment-identity invariant — the artifact your gate validates must be byte-linked to the artifact the game loads (added 2026-07-29, SUB_BRZ S42): a native-contract gate that consumes a separate ODOL stage says NOTHING about the deployed PBO. -packonly copies whatever format sits in the compilable tree, and pairing MLOD↔ODOL stage files by homonymous path is not closure. Measured case: the gate validated the ODOL stage doors (~2.1 MB each) while the live PBO carried the source MLOD (8.7/8.1 MB, byte-identical to the MLOD stage) — nothing in the pipeline linked the two artifacts, and the mismatch surfaced only during post-crash analysis. Rule, before ANY deploy of a gated artifact: list the PBO, extract or slice every .p3d entry by offset, and require (a) the expected signature (ODOL where the contract is ODOL) and (b) SHA-256 equal to the approved stage — fail on wrong signature, missing entry, casefold/slash collision, or hash mismatch. Corollary: record the mounted PBO's SHA-256 contemporaneously at every runtime launch; without that seal, no post-crash analysis can prove which artifact the process actually loaded.

Native-crash fingerprint — compare CODE BYTES, not fault addresses (added 2026-07-30, SUB_BRZ S42): the RPT's Fault address is not an identifier — ASLR moves the module base, so the same code site prints different addresses across runs and reads as "different bugs". What identifies the detection site is Prev. code bytes + Fault code bytes (16 bytes each) plus the engine Version. Measured case: six C0000374 crashes split across two addresses (3523D5F9 and 57B5D5F9) had byte-identical code bytes and identical low 16 bits — one code site, one engine build. Reading them by address had produced the (false, and formally rejected) claim that the ODOL-era and MLOD-era crashes were unrelated; the byte comparison also killed the hypothesis that the artifact FORMAT was the discriminating variable, since the crash appeared in both. Report "same/different detection signature", never "same/different bug": C0000374 is detected on free, not where the out-of-bounds write happened, so an identical signature still does not prove identical cause.

ITERATION BUDGET — check this BEFORE opening a front (added 2026-07-27)

Measured over one month of vehicle work (125 session handoffs: MercedesAMGLF 46, SUB_BRZ 38, LFHeli 37): 26 sessions ended in STOP/BLOCKED/FALSIFIED against 24 green. The three failure modes below are process gates, not config facts — and each burned dozens of sessions on a different project before being named. They cost more iterations than any invariant below.

  1. Count the cycles spent on one objective, and make the count block. DZ-R5 already says "after 2 rebuilds without progress, STOP and change strategy"; it does not bite because nobody tracks N. Carry ciclos_en_este_objetivo: N in the HANDOFF LIVE-STATE, and at N>=3 the re-entry gate must pick one of three exits before touching code: change strategy, escalate to a human, or drop the objective. Evidence: MercedesAMGLF Task 9 reached v64 across 29 sessions, and what finally closed it was shifting 28 proxy triangles by -0.240 m in X.

  2. Keep the test harness OUT of the product task. Task 9 mixed "align the Mercedes proxies" (product) with "build an automated live verifier" (daemon, broker, lease, preflight, camera contract, JPEG SHA) — the harness silently took the bulk of the time and became the project. If a product task is blocked TWICE by the harness, stop the product and decide explicitly: scope the harness as its own project, or run that one test by hand and move on.

  3. When the judge is the eye, offline analysis gets ONE round. Green offline gates do not predict the live result. A Task 9 build reported 263/263 global + verifier 35/35 hard PASS plus every SHA, and then: Visual: FAIL of evidence, client hung. The LFHeli "seis frentes" night spent ~25 Codex sessions to produce refutations only (binding intact, re-bake refuted, clock and interpolation refuted) and concluded the discriminator was the user's own screenshot. If the discriminator is visual or feel, spend one offline round ruling out cheap causes, then go to the game — extra offline rounds postpone the verdict, they do not replace it.

Corollary already paid twice: SUB_BRZ needed 30+ sessions to conclude that a decimated rip mesh cannot reach game-ready quality (pivot to a human artist), and the LFHeli HH-60G reached the same verdict independently weeks later. A visual red that survives N rounds is a signal to escalate to a human, not to run round N+1.

REDIRECT CAMBIO-1: familia B → ../rip-vehicle-import/cookbooks/family-b/radial-puerta-ausente.md.

INV-ALIGN — a faithful import does NOT mean the pieces line up (added 2026-07-27)

A bit-exact import and correctly-fitting pieces are different properties. When a model is assembled by script, each piece is placed at a procedurally computed anchor (centroid, bbox centre, a reference point). If that anchor diverges from the real reference, the pieces sit wrong while every import check passes. Paid three times (MercedesAMGLF, SUB_BRZ, LFHeli):

  • MercedesAMGLF: the 28 wheel proxies sat -0.240 m in X from their Memory hubs. Fix: measure the offset with py3d and shift only those triangles (max error 0.0 m afterwards).
  • LFHeli OH-1: import fidelity 12/12 at max 3.2e-7 m, and the SEATS procedural anchor still diverged 583 mm from the source mesh.

Gate before declaring an assembly good: for every proxy, measure its anchor against the real reference (the matching Memory point, or the source mesh position) and assert the delta is ~0. Import fidelity, digest equality and structural diffs do NOT catch this — they compare the piece to itself, not to where it belongs.

Do NOT "fix" it by adjusting the pipeline transform. That is the intuitive move and it is a false fix: it displaces everything else that was already correct. Correct the individual anchor.

False friend — autocenter=0 is not this bug. It is a real and separate requirement (the engine recenters LODs by bbox without it, displacing collision ~20-25 cm from visual; see dayz-model-pipeline and dayz-p3d-audit), but it does not cause or cure anchor divergence. MercedesAMGLF spent a full RCA on that hypothesis and closed it as unproven: adding autocenter=0 to all 18 proxy LODs changed nothing geometric, and the protrusions and see-through were already present in the earlier baseline. Do not re-run that investigation.

TEXTURE OR MESH? — discriminate BEFORE auditing the texture pipeline (added 2026-07-27)

"The textures look wrong" on an imported vehicle is a symptom with at least three unrelated causes, and auditing the texture pipeline is the most expensive way to find out which. Run this ladder first; each rung is minutes, the full audit is sessions.

  1. Audit the UVs, not the textures. Exact SAT overlap + island count + density (uv-clean-atlas, uv_audit.py). Overlap > 0, atomised islands (2-4 faces each) or wildly uneven density means the problem is upstream of any .rvmat — no material wiring fixes it.

    No-overlap means an exact SAT=0 (SP-214, added 2026-08-31). A Monte-Carlo estimate has a measured nonzero floor, so its displayed 0% is not proof. Route vehicle UV repair to uv-clean-atlas and use its current chart/SLIM/fold/SAT and semantic-packing contract; PartUV is not the vehicle default. One anti-fold round is a cost target, not a correctness exception: run the same exact SAT detector after every round, continue when it is nonzero, and block any mesh that does not reach zero. This pack owns the routing and acceptance gate, not a second copy of that implementation.

  2. Ask where the mesh came from. A decimated rip cannot yield game-ready UVs: the cause is topology, not the unwrap algorithm or the texture route. The fix is retopo (bake the high-poly onto a clean low-poly), and no amount of pipeline correction substitutes for it.

  3. Render the same material on a known-good control. If a vanilla-derived control looks right with the same route, the pipeline is exonerated and the defect is in this asset.

  4. Only now audit .rvmat stages, _co/_nohq/_smdi paths, .paa conversion and hiddenSelections — see dayz-texture-pipeline, and its references/vehicle-materials-and-color-variants.md for the vehicle-specific shader choice (Super where faces carry a _co, NormalMapSpecularMap for constant-colour parts, tiled detail for overlapping or un-baked UVs).

Cost rule: if two rounds of pipeline auditing come back clean, "the texture is applied wrong" is REFUTED — change branch instead of running a third. An audit that exonerates the pipeline is evidence about where the bug is NOT, and repeating it produces no new information.

Evidence for the rule: on LFHeli OH-1, F6 ran a double audit that exculpated the entire offline chain (UVs identical across blend/MLOD/ODOL, old PAA identical to new, rvmat uvSource=tex, zero hiddenSelections, no faces inside the watermark box) — the textures were never the problem. On SUB_BRZ the documented root cause was the decimated rip's topology, reached after six rounds and a pivot to a human artist. In both cases the ladder above would have branched away on rung 1 or 2.

Rip-specific rung 0 — a "paint" material in the source has NO diffuse; never substitute a livery sheet (added 2026-07-27)

Before running the ladder on a ripped vehicle (GTA, ripped racing-game, any game rip), census which source material each face carries and what the import mapped it to. Game rips colour the body through the host game's paint system, not through a texture: in GTA those materials are vehicle_generic_smallspecmap* / "primary" and their diffuse is legitimately absent or a zero-byte file. An importer that needs some texture per face will silently fall back to whatever is available — and if what is available is the model's livery/decal sheet, the result is signs, numbers and camo stamped across the whole fuselage. The symptom reads as "the UVs are wrong" or "textures are crossed", and both readings send you down the expensive branch.

  • Discriminator (minutes): count faces per (source material -> texture) pair. A single texture carrying a large share of the model, especially a decal/livery atlas, is the signature. Cross-check the source file size on disk: a 0-byte or missing DDS proves the fallback fired.
  • Fix direction: those faces want a flat colour _co (the host game's paint slot has no DayZ equivalent), not the decal sheet and not a re-unwrap. Their UVs were never meant to address a texture, so no UV work is involved.
  • Measured evidence. LFHeli HH-60G (2026-07-27): hh60g_sign_1_co.paa — a 22.4 MB livery sheet — carried 23,716 / 35,385 faces (67%), of which 8,695 (24.6% of the model) were pure fallback (7,568 of the body's [PRIMARY] material plus all four doors), with vehicle_generic_smallspecmap3.dds confirmed 0 bytes on disk. The user's independent report was "things and colours drawn where they don't belong". The OH-1 line hit the same class separately (rotor blades wearing camo and a registration number) after the PAA route had been exonerated.
  • Status honesty: the diagnosis is measured; the flat-colour fix was not yet confirmed in-game at the time of writing — treat the fix direction as unverified until a cycle closes it.

A mesh DRAWN in the wrong place is invisible to every script transform — qualify the probe first (SP-138, added 2026-07-29)

When the user reports that a vehicle draws somewhere the entity is not — body on the ground while the vehicle climbs, a part left behind, a piece floating — the reflex is to instrument the transforms and compare. That reflex has already been paid for and it returns nothing.

Measured on LFHeli OH-1 over a 26 m climb: root position, GetRenderTransform, both bones, the physics body (dBodyGetWorldTransform) and the pilot's world position agreed digit-for-digit while the hull was drawn on the ground. Recorded in that project's own code at LFHeliCore\scripts\4_world\LFHeli\LFHeli_Base.c:3106-3113. The consequence is the general rule:

Any probe derived from the entity root inherits the root's blind spot. A drawing-position defect lives downstream of every transform Enforce exposes, so no amount of transform logging can see it.

What does work — a marker the engine projects from a WORLD point without consulting the entity's transform, read off a screenshot: Debug.DrawSphere at GetPosition() plus a ladder of rungs at fixed spacing below it. The rung spacing is what makes the defect quantitative (gap in metres = rungs × spacing) instead of an eye report. Reference implementation: LFHeli_Base.c:3114-3128.

Probe qualification, mandatory before any number adjudicates: read the magnitude in the regime where the drawing is CORRECT (on the ground) and in the regime where it is WRONG (airborne). If it does not change while the eye sees a difference, that magnitude is blind — stop using it. A delta of exactly 0.000 is a tautology signal, not precision.

The companion trap, on the inference side. Over flat ground, "the visual Y is frozen at its take-off value" and "the transform is computed from a ground-referenced magnitude" produce the same two signatures: gap grows with height, gap returns to zero on landing. Arithmetic, confirmed by external review. So the flat-ground signature must never be used to EXCLUDE a family of causes — it identifies nothing. Discriminate by changing the regime, not by reasoning:

  • horizontal traverse at constant world Y over terrain ≥15 m lower — ground-referenced tracks the terrain, frozen-Y does not;

  • descend without touching down and climb again — a replication/state split shows hysteresis, a pure function of height does not.

  • Status honesty: the blind-spot measurement and the flat-ground ambiguity are verified; the two discriminant manoeuvres above were designed on 2026-07-29 and had not been run in-game at the time of writing. Treat them as an untested procedure until a cycle closes it.

REDIRECT CAMBIO-1: familia B → ../rip-vehicle-import/cookbooks/family-b/get-in-ausente.md.

INVARIANTS YOU WILL HIT — preflight checklist (read BEFORE authoring, not after the in-game fail)

These recur on every vehicle. They were each won the hard way on one project and then re-derived from scratch over dozens of iterations on the next (LFQuad → SUB_BRZ → MercedesAMGLF), because they were not promoted to this checklist in time. Check them up front; full detail behind each pointer.

BEFORE copying a donor — verify WHERE the engine is, not how many seats it has (SP-205, origin 2026-08-08, SUB_WRXSTI T8). sedan_02.c looks like the vanilla 4-seat sedan control and is rear-engine: it closes CarBattery and SparkPlug behind the trunk and the CarRadiator behind the hood (scripts\4_world\entities\vehicles\inheritedcars\sedan_02.c:142-160), and gates CanDisplayCargo on the hood (sedan_02.c:203). Porting that to a front-engine car leaves the engine parts behind the wrong lid — a bug that only shows in-game. The correct control for front-engine + 4 seats is civiliansedan.c. Rule: before copying a donor, verify WHERE the engine lives, not only how many seats it has. Rear seat on that 4-seater (measured vs civiliansedan_mlod.p3d; in-game confirmation pending WRX live test #1): Memory pos_cargo1(_dir), pos_cargo2(_dir), seat_con_1_2, seat_con_2_2; GetSeatAnimationType cases 2/3 → VEHICLESEAT_PASSENGER_L / VEHICLESEAT_PASSENGER_R and CrewCanGetThrough cases 2/3 (civiliansedan.c:119-131, :226-256) — the base returns false (transport.c:727-733). Distinct crew_cargo proxy index and seat-box lead: invariants 28 and 30.

Four-seat completion rider (SP-205, added 2026-08-31). Rear seats also need seat_cargo1/2 boxes dual-tagged to their exact componentNN in Geometry and ViewGeometry; crewcargo1/2 proxies in ViewGeometry and FireGeometry using one crew_cargo path with distinct instance indexes; Crew.Cargo1/2 entries with actionSel, proxyPos, getInPos, and getInDir; and the two passenger cases in both GetSeatAnimationType and CrewCanGetThrough. CarScript already maps all four seat_* selections to seat_con_1_1, seat_con_2_1, seat_con_1_2, and seat_con_2_2 (carscript.c:2141-2156), so an override of GetDoorConditionPointFromSelection that only repeats that table is dead code.

That base table is the LEFT-HAND-DRIVE mapping (SP-426, added 2026-10-01, WRX STI B-01). [EXACT][CLAIM-GETIN-LHD-BASE-TABLE] On a right-hand-drive car whose memory points are named by physical side (driver door = seat_con_2_1), the base table pairs every seat with the FAR door: each seat is evaluated against the opposite door and CanReachSeatFromDoors fails the 1.0 m plan-distance test from the correct one, so no get-in appears with names, points and componentNN all correct. Signature in query_get_in_condition: unreachable from the correct door and available from the opposite one. Override it crossing the sides (driver->seat_con_2_1, codriver->seat_con_1_1, cargo1->seat_con_2_2, cargo2->seat_con_1_2), as vanilla does for its own layouts (OffroadHatchback.c:364-379, Van_01.c:269-287, DayZ 1.30.164014 Exp). On RHD cross GetSeatAnimationType for the rear seats too (vanilla pairs L with the physical left side, CivilianSedan.c:119-134); the driver keeps VEHICLESEAT_DRIVER, whose get-in animation will look LHD-sided - an engine limit, not a mod bug.

BEFORE any numbered gate below — calibrate the gate itself (SP-132, added 2026-07-29; LFHeli HH-60G + OH-1). Two rules in this file already cover this — §"Calibration + scope (so a gate is trustworthy, not a false green)" and §METHOD habit 2 ("never close a phase on a tautological gate"). Both were live and neither fired across four gates in one week, because both read as being about car render/winding gates. They are not: they govern every item below, and capacity, parity, telemetry and policy gates alike. The four shapes, so they are recognizable on sight:

  • A threshold copied instead of measured. hh60g_assemble.py:29 carried RESOLVED_LIMIT = 65535; the engine's real resolved-vertex ceiling on that model is 46.133 (42% lower), so the gate passed models binarize rejects with Too many vertices. The cap is model-dependent (×1.4205 on HH-60G, ×1.46 on OH-1) — never copy it between models, re-measure with binarize.
  • A policy that is a mathematical no-op. uniform and keepdot were geometrically inert (dot(-cross,-n) == dot(cross,n)); four variants were built on them before the identity was checked.
  • A gate that shares the producer's frame. OH-1 seat parity compared under pure translation — the same assumption the sub-model was generated with. delta 0.000 proved producer and gate agreed, not that either was right.
  • A field read selectively. attachment_count=2 was signed PASS because that one field matched expectation; the regression sat in the same output 20 min before the user hit it in-game.

A fifth shape, and it fakes an ABSENCE rather than a green (added 2026-08-19, SUB_WRXSTI): matching a LOD by its nominal resolution. LOD resolutions on disk are not the round numbers the docs use — FireGeometry reads 6999999976046592, not 7e15. A bucket keyed on round(7e15), or any == against the nominal, comes back EMPTY, and a gate built on it reports "the selection is absent" about something that is present. Measured cost: a cross-check of the config's componentNames[] against the .p3d reported all six damage zones missing, and they were all there in FireGeo — the report reached the user as a red before the control against vanilla caught it. Always match with a RELATIVE tolerance (abs(res - nominal) <= nominal * 1e-3), and note the direction of the failure: this one does not show up as a suspicious green you might question, but as a confident red about missing data, which is the shape people act on fastest.

One minute, before trusting any gate: (a) run it against a KNOWN-BAD case — green on a known-bad means the bug is the gate, not the artifact; (b) name in one line the assumption the artifact was BUILT with, and confirm the gate does not reuse it. A gate that has only ever been seen green is unmeasured, not verified.

  • Gates that stay red stop being sensors (LL-244). If a known failure is deliberately accepted, pin its exact signature — entities and measurements — and report KNOWN_RED separately from NEW_RED. Any drift, including an improvement, is new evidence because the pin describes a state, not a ceiling. A gate that returns the same verdict for N runs without making a new red distinguishable has become a constant.
  1. Gate #0 — mesh + UV health BEFORE anything else (SP-052). For ANY imported model (ripped or not), the FIRST step is a mesh+UV audit, because a broken mesh produces false downstream diagnostics (a "double wall" that was mirrored UV on a single wall; a "convexity" issue that was open design-boundary edges). (a) audit mesh_health.py (boundary / non-manifold / dup-verts / dup-faces / winding per piece) + uv_audit.py (islands / overlap / density / mirrored); (b) sanitize with an ADAPTIVE weld (ladder 0.5→0.01 mm; gates: NM must NOT grow + UVs preserved — a blind 0.5 mm weld FUSES legitimately-touching pieces and CREATES non-manifold); (c) unwrap BY SECTIONS per AI/20_Knowledge/uv-mapping-dayz.md §0.5/§5; (d) gates: overlap ~0% for bake, uniform density, atlas legible by a human, checker render with SMOOTH shading (flat shading + grid raster read as a false "low-poly/blurry"). Tools in LFQuad_dev\tools\ (mesh_health.py, mesh_sanitize.py bpy-headless; uv_audit.py py3d) — portable to any vehicle. → AI/20_Knowledge/uv-mapping-dayz.md, winding #10(f).

0b. Gate #0b — when LOD1/2 come from DECIMATE instead of an artist, audit the modifier's OUTPUT (added 2026-08-06, MercedesAMGLF). No artist is now the normal case, so per-object COLLAPSE + planar DISSOLVE is the default LOD path. The modifier leaves loose vertices — vertices no triangle indexes. Every count-based gate stays green (triangles, resolved, identity, monotonicity, frame parity), and the streams ship broken: a per-vertex extract loop driven by mesh.loops never reaches a loose vertex, so whatever the array was initialised to is what gets written. Initialised to (0,0,0) → zero-length normals in NORMAL.bin. Measured: source had 0 loose in 224.700 verts; after decimation, 786 at LOD1 and 4.024 at LOD2 — plus 16 normals of length down to 3e-6 that a > 0 usability test happily passed. Three checks close the whole class:

  • delete loose verts after every decimate (bmesh.ops.delete(bm, geom=[v for v in bm.verts if not v.link_faces], context="VERTS")). NOT one of the three forbidden calls — it removes only geometry no triangle references. Skip the bmesh round-trip entirely when there is nothing to delete, so must-keep pieces never pass through it.
  • assert every emitted normal is unit length and normalise on emission. The real cut is 0.5, not 0: a unit normal survives float32 within ~1e-7 of 1.0, so anything shorter is a cancelled sum. Substitute from NEIGHBOUR normals, never from poly.normal — if the import transform negates an axis it is a mirror, and the winding-derived normal points the opposite way (see #10 winding).
  • assert every index falls inside its own piece's vertex range. No count gate sees an out-of-range index; it explodes at assembly.

Gate corollary — a drift number that mixes classes is not a measurement. Split survivors into INTACT (incident-triangle set identical to LOD0 → drift MUST be ~0; any drift here is a real defect), RECOMPUTED (topology changed → drift is legitimate) and DEGENERATE (counted, never averaged in). Mixed together, a p99 of 90.00° hid a real defect AND hid that untouched geometry was clean at ≤0.26°. An exact round angle is a fingerprint, not a result: 90.00 is acos(0) = a null vector on one side. Match survivors on a ~1e-4 m grid, not 1e-6 — positions are written as float32 and a 1e-6 grid drops about a third of the untouched vertices at random, which is why the samples looked tiny.

Position is a candidate bucket, not identity (SP-163, added 2026-08-31). The grid above may find candidates, but a seam can place distinct normal/UV corners at the same coordinate. Pair intact data by a carried source vertex/corner and primitive ID. If that identity was lost and a bucket is non-unique, return INCONCLUSIVE, not PASS or FAIL. Run source against itself through the same matcher and require approximately zero residual before trusting any product comparison.

  1. "Get in" radial needs a real SCRIPT CLASS, not a bare config class. A class <Mod>: CarScript with no .c runs as bare CarScript, which inherits Transport.CrewCanGetThrough() → false (scripts/3_game/vehicles/transport.c:727) → the get-in action is filtered out and never appears, no bone/proxy/componentNN fix helps. Author <Mod>_Base extends CarScript overriding CrewCanGetThrough + GetAnimInstance + GetSeatAnimationType (+ CfgMods worldScriptModule). Verified in-game LFQuad D34 + MercedesAMGLF + SUB_BRZ. → vehicle-structural-parity.md "Crew get-in".

  2. The script module must actually LOAD or the class never binds — silent. CfgMods files[] with back-slashes / a .p3d path that resolves to *.p3d.p3d → module not loaded → no get-in, no error. → SKILL.md §"Binding del script", rip-import.md.

  3. REDIRECT CAMBIO-1 (familia B): ../rip-vehicle-import/cookbooks/family-b/wheelpresent-0.md.

  4. Seats and wheel hubs must be componentNN-tagged (dual-tag) — AND each seat component must be raycast-collidable. The engine enumerates collision components only by componentNN; a seat/hub that is a standalone island (0% componentNN overlap) is invisible → spawn blocker / no seat. Tag the SAME faces with a componentNN too. Then the ViewGeo seat cube must be INWARD-wound with every point flag = 0x0000003F: a py3d box left outward + flags 0 passes every offline shape/winding/dual-tag check yet RaycastRV(ObjIntersectView) does NOT hit it → the get-in cursor falls through to component0 → the driver "works" by fallback but the CODRIVER never resolves. The decisive, in-game-confirmed copilot blocker on BOTH SUB_BRZ (s9) and MercedesAMGLF (s12, headless hit=1 comp=6 crewIdx=1). The s9 patch changed winding and flags together and never isolated the flag as cause; the safe rule is to copy the sealed vanilla control's convention, never py3d's default. → "componentNN DUAL-TAG" + "CRITICAL EXTENSION 2026-06-28" (vehicle-structural-parity.md). positive control (LFQuad/Croco), never trust py3d's default. → "componentNN DUAL-TAG" + "CRITICAL EXTENSION 2026-06-28" (vehicle-structural-parity.md). Resolve the seat's componentNN; never hardcode its number (SP-079, added 2026-08-31). Assembly order moves component numbers. For every seat_* selection, compare both point and face membership against each componentNN in the same LOD and require exactly one exact 1:1 match. Zero or multiple matches is a hard failure. Re-run the equality after any transform or transfer so the seat/component dual tag cannot be lost silently.

  5. Crew/wheel proxies must exist in BOTH ViewGeometry AND FireGeometry; the proxy triangle uses the engine identity frame (R=((-1,0,0),(0,0,1),(0,1,0)), model-space), NOT py3d rotation=None. → parity + rip-import.

5b. wheelHub names a GEOMETRY/MEMORY selection, not the visual wheel_X_Y — and the visual wheel_X_Y IS the proxy's own triangle under a second name (added 2026-08-07, MercedesAMGLF). class Wheels { class Left { wheelHub="wheel_1_1_damper_land"; } } resolves to a selection that lives in Geometry 1e13 (a ~0.20 m hub box) and Memory 1e15 (1 point) — vanilla DZ/vehicles/wheeled/config.cpp:297. It is NOT the wheel_X_Y selection on the visual LOD. Measured on two independent cars: that visual companion has the SAME centroid and the SAME bbox as its proxy:<path>.NNN selection, i.e. it is one triangle carrying two names (vanilla shows it as 1 face too). Consequences, each of which cost a session:

  • A gate of the form "proxy anchor -> companion centroid == 0" measures nothing about the wheel. It computes |anchor - centroid| of a single triangle, which is a pure function of that triangle's SIZE. Measured on the same model: 0.131757 m for a 0.339x0.204 triangle vs 0.000745 m for py3d's canonical scale=0.001. Against the real hub both artifacts read 0.000000 m. Point the gate at wheel_X_Y_damper_land.
  • Proxy triangle size bakes NO scale. The frame is derived from UNIT vectors (py3d/__init__.py:174-179, port of proxy_frame.derive_frame), and vanilla's binarised proxy matrix is orthonormal (|aside|=|up|=|dir|=1.000000 on civiliansedan.p3d ODOL v54). A 1 mm canonical triangle and a 0.34 m one place the same wheel; SUB_BRZ shipped on scale=0.001. Do not "fix" a small triangle.
  • Vanilla reference, if you need one: civiliansedan puts the hub 0.106334-0.107333 m INBOARD of the wheel proxy, the whole delta in X. Hub coincident with the proxy (0 m) is tighter than vanilla, not a defect.
  • Before matching a magnitude to another open debt, decompose it by AXIS. 0.1318 m was read as "the wheel drawn 13 cm off its simulated position" because it resembled a 0.136 m track debt; the 0.1318 lives entirely in Y/Z (dX = 0.000000 exactly on all four wheels) and the track debt lives entirely in X. Orthogonal — the resemblance was numerology.
  • Read a gate's PRODUCT verdict from a run WITHOUT its negative fixtures, then repeat it on the known-good artifact as a control. Run with --negative-fixture and the fixture's deliberate reds (yaw-180 families, stripped companions) interleave with the product rows and read as product failures.

5c. Owner-only driving hitch: audit hub-ground clearance and remove LandContact before blaming scripts (SP-355, added 2026-08-31). The driver owns the physics solve while a passenger interpolates, so smooth passenger motion does not clear collision geometry. Working DayZ car controls use the tyre's PhysX sphere plus a small wheel_*_damper_land Geometry box (about 0.20×0.22×0.22 m) and no LandContact LOD. Require hub_y - mounted_tyre_radius to exceed both Suspension.travelMaxUp and `tra

Files (dayz-modding-knowledge-pack)
  • history
    • cambio-1-superseded-family-b-rules.md 11.6 KB
      Loading…
    • pre-cambio-0-gate-ladder.md 6 KB
      Loading…
  • references
    • animation-sign-and-axis.md 6.3 KB
      Loading…
    • audit_getin_wheels.py 4.9 KB
      Loading…
    • binarize-vertex-budget.md 10.4 KB
      Loading…
    • build-packaging-and-debug.md 5 KB
      Loading…
    • check_odol_bone_refs.py 2.6 KB
      Loading…
    • dayz-1-30-vehicles.md 29.3 KB
      Loading…
    • decimation-libraries.md 11.3 KB
      Loading…
    • doors-and-panel-edges.md 13.3 KB
      Loading…
    • emissive-leds-and-dynamic-lights.md 13.6 KB
      Loading…
    • external-obj-import.md 15.3 KB
      Loading…
    • gate_odol_sections.py 8.2 KB
      Loading…
    • gauge-needles.md 70.8 KB
      Loading…
    • get-in-actions-and-proxy-pose.md 12.1 KB
      Loading…
    • gunner-shoot-from-vehicle.md 9.2 KB
      Loading…
    • legacy-gate-ladder.md 5 KB
      Loading…
    • materials-and-selections.md 5.4 KB
      Loading…
    • network-physics-ownership.md 13.9 KB
      Loading…
    • placement-and-autocenter.md 5.1 KB
      Loading…
    • png-to-paa-encoding.md 10.4 KB
      Loading…
    • process-gates.md 11.5 KB
      Loading…
    • proxies-and-get-in.md 39.4 KB
      Loading…
    • rip-geometry-and-winding.md 15.7 KB
      Loading…
    • rip-import.md 85.6 KB
      Loading…
    • vehicle-config-and-modelcfg.md 67.1 KB
      Loading…
    • vehicle-structural-parity.md 97 KB
      Loading…
    • vehicle-types-boat-truck.md 15.1 KB
      Loading…
    • visual-gates-and-winding.md 35 KB
      Loading…
    • visual-metrics-and-viewer.md 8.4 KB
      Loading…
  • SKILL.md 183.1 KB
    ---
    name: dayz-vehicles
    description: >
      Author, import, debug and ship DayZ ground/water vehicles: CarScript cars,
      trucks, quads/ATVs and motorbikes plus Boat-based watercraft. Covers the
      config.cpp/model.cfg contract, drivetrain/engine/suspension, Axles/Wheels,
      Crew/get-in/actions, AnimationSources, structural parity, LODs, wheel/crew
      proxies, View/Fire/Geometry, placement, winding and dedicated-server
      packaging failures. Use for car/truck/quad/ATV/moto/boat, drivable vehicle,
      CarScript, civiliansedan parity, importing Blender/OBJ/ripped racing-game vehicles,
      "get-in broken", wheels missing/not steering/spinning backwards, white or
      untextured vehicle, vehicle build/pack/binarize or proxy pose. Always invoke
      before authoring/debugging a vehicle entity, config.cpp or model.cfg. Use
      dayz-aviation for flight and dayz-model-pipeline for generic assembly.
    ---
    
    # DayZ Ground Vehicle Modding
    
    ## Family selector — day-0 critical path (CAMBIO-2)
    
    1. Identify family by asset origin and architecture; not by symptom.
    2. If asset is a new Grub source-game car, proxy-split and with moving parts, use **family B**.
    3. Family B → open `../rip-vehicle-import/SKILL.md` and follow only its adapter, golden, and allowlist.
    4. That adapter opens asset `asset-contract.json` as third and final day-0 file; schema, Blender export, and primitive are machine inputs, not additional documents maintained by agent.
    5. If a frozen plan/runbook already exists for asset, follow that contract in flight; do not migrate it here.
    6. If no row applies or family adapter is missing: **STOP**. Do not improvise from this atlas.
    
    | Input signal | Adapter | Day-0 files |
    |---|---|---|
    | source-game Grub, new car, proxy-split + doors/moving parts | Family B | This router → `../rip-vehicle-import/SKILL.md` → `<asset>\asset-contract.json` |
    | Asset already in flight | Frozen project runbook | Whichever cited in its current plan |
    | Other family or insufficient data | None | **STOP** |
    
    Rest of this body is diagnostic atlas and **does not form part of day-0 critical path**. Only consulted when adapter or symptom links a specific section.
    
    **Preflight before re-importing / re-assembling any piece** (SP-193): first measure FACTORY FACE-LEVEL completeness of piece against already packed model. A "rotten" mesh is usually complete and defect lies elsewhere; see SP-193 at end of this atlas.
    
    ---
    
    Drivable wheeled vehicles in DayZ extend **`CarScript`** — there is no generic "vehicle" base, and
    you do not build one from `EntityAI` or `Inventory_Base`. `CarScript` gives wheels, seats, sound
    integration and the damage system for free; you configure the rest. This skill is the
    vehicle-specific layer on top of `dayz-model-pipeline` (generic geometry / LODs / textures) and is
    the ground counterpart to `dayz-aviation` (anything that flies belongs there).
    
    Scope: cars, trucks, quads/ATVs, motorbikes — anything wheeled and drivable, modeled from scratch
    or imported (Blender / OBJ) from another game.
    
    > REDIRECT CAMBIO-1: header family selector and chosen adapter are the sole day-0 path.
    
    **Vehicle type matrix (invariant):** `Car` and `Boat` are SIBLINGS, both directly under `Transport`
    (`car.c:87` / `boat.c:28`) — NOT parent/child. `Transport` owns crew/get-in/flip/fuel; `Car` owns
    wheels/brakes/`CarFluid`; `Boat` owns propeller/buoyancy/`BoatFluid`(fuel-only). A truck is a plain
    `CarScript` config with 3 axles + double wheels, NOT a new class. Boats, truck double-wheels, ATV and
    the motorbike gap → `references/vehicle-types-boat-truck.md` (historical: since DayZ 1.30 Exp motorbikes are a vanilla family, see the `dayz-motorbikes` skill).
    
    **DayZ 1.30 Exp (build 1.30.164014) changes the matrix (added 2026-09-16, source_verified on the extraction at `E:\DayZ-Exp-Extract\1.30.164014\exp`):**
    - [EXACT] A third sibling exists: `class Motorbike: Transport` with `simulation = "motorbike"` (`dta\bin\config.cpp:1054-1059`; engine proto `3_Game\Vehicles\Motorbike.c:30`), scripted by `MotorbikeScript` (`4_World\Entities\Vehicles\MotorbikeScript.c:53`), which does NOT inherit `CarScript`. Two-wheel work goes to the `dayz-motorbikes` skill; this atlas stays for cars, trucks, quads and boats, plus the ViewPilot and the first-person view of any family (SP-248, SP-249 and `references/proxies-and-get-in.md`).
    - [EXACT] `CarScript` was refactored around components owned by `Transport`: `VehicleLightsComponent` (`CarScript.c:187, 229-233`), `VehicleHornComponent` and `VehicleVFXComponent` (`CarScript.c:189, 343-368`). The old overrides are marked `[Obsolete]`: `CreateRearLight`/`CreateFrontLight` (`CarScript.c:2937-2945`), `IsVitalCarBattery`/`IsVitalTruckBattery` (`CarScript.c:2811-2818`, replaced by `Transport.NeedElectricitySourceDevice()` at `Transport.c:830`), `CarPartsHealthCheck` (`CarScript.c:2976-2979`, replaced by `Transport.PartsHealthCheck()` at `Transport.c:440` / override at `CarScript.c:2121`). The horn actions are `ActionVehicleHornShort/Long` (`CarScript.c:2319-2320`). Contact data is `VehicleContactData.m_Impulse` (`Transport.c:1153-1157`, used at `CarScript.c:1291-1292`). Wheel classes gained `tyreRoughness`, `tyreLongitudinalFriction`, `tyreLateralFriction` (`DZ\vehicles\wheeled\config.cpp:749-751`; absent in 1.29).
    - [DESIGN] A 1.29 car mod that overrides any of the obsolete members compiles against the attribute but its override is no longer called; re-verify each override against the 1.30 `CarScript.c` before shipping. Migration checklist and the "1.29 statements that are now false" table: `dayz-motorbikes/references/vehicle-1.30-refactor.md`. Sections of this atlas that predate 1.30 (lights via `CarLightBase`, `CarPartsHealthCheck`, `data[0].impulse`) describe 1.29 and are not yet rewritten.
    - [EXACT] `Transport.GetWeightCoef()` default `1.0` (`Transport.c:478-481`); vanilla `Truck_01_Base` `2.0` (`Truck_01_Base.c:120-123`), `Offroad_02` `1.5` (`Offroad_02.c:133-136`), `Motorbike_01` `0.5` (`Motorbike_01.c:84-87`), `Motorbike_02` `0.75` (`Motorbike_02.c:74-77`). Digest M's claim that `MotorbikeScript.c:661` *returns* `0.25` is false: that line *uses* `GetWeightCoef()`, it does not define it.
    - [EXACT] `CarScript.HasGangedTailAndBrakeLights()` default `false` (`CarScript.c:468-471`); `true` on Offroad_02 (`Offroad_02.c:68-71`) so reverse/position swaps do not kill the brake selection (`CarScript.c:2061-2094`).
    - [EXACT] Vehicle light toggle is `UAToggleVehicleLights` / `ToggleVehicleLightsActionInput` (`ActionInput.c:849-855`, `ActionSwitchLights.c:16-18`); `UAToggleHeadlight` is headgear (`ActionInput.c:776-782`). `ForceUpdateLightsStart`/`End` are `[Obsolete("1.30: no replacement")]` (`CarScript.c:2805-2809`).
    - [EXACT] Get-in calls `CanGetIn()` then `IsAreaAtDoorFreeDiag` and **does** call `CanReachSeatFromDoors(..., 1.0)` from script (`ActionGetInTransport.c:36,54,64`).
    - [EXACT][CLAIM-VEH-130-PORT-BLOCKERS] Two 1.30 changes break a 1.29 car outright, beyond its looks. `CarScript.ToggleHeadlights()` was deleted without an `[Obsolete]` stub (1.29 `CarScript.c:2088`), so an `override` of it stops the `4_World` compile. A car that does not override the three electricity-source methods reports `NO_BATTERY` with a battery attached (1.30.164200 `transport.c:835-838`, `carscript.c:1768-1774`): no engine start and no light switch. Two more showed up in game on 1.30.164200: a backslash script module path loaded none of the LFQuad's scripts, silently, and every wheel class without a `<class>_Ruined` twin raised a script error when its first wheel was created (1.30.164200 `inventoryitemtype.c:17-46`). Detail and the override check: `references/dayz-1-30-vehicles.md` (table and the path claim below it, section 2, section 11, checklist items 0 and 11-13). Seen porting the LFQuad to 1.30.164200 (2026-10-26).
    - Lights, horn, fluids, boats, wreck cisterns-as-Well, `speedGeoms`, and the motorbike 3PP camera pointer: `references/dayz-1-30-vehicles.md`. Motorbike camera ID 32 lives in `dayz-motorbikes/references/rider-animation.md`.
    
    
    **Per-tick event asymmetry (invariant, added 2026-07-29, LFHeli OH-1):** `CarScript` has NO
    `EOnSimulate` — its ctor registers only `POSTSIMULATE`/`POSTFRAME` (`carscript.c:207-208`) and ALL
    its per-tick work (engine, fluids, part health, fuel/engine auto-stop, exhaust and wheel FX, much of
    it gated `IsServerOrOwner()` — vanilla runs it on the owner client, but in mods replica-dependent checks (`CheckVitalItem`) must stay server-only (LL-305)) lives in `EOnPostSimulate`
    (`carscript.c:807`). `BoatScript` overrides BOTH (`boatscript.c:355` and `:393`). Consequences on a
    car: hook `EOnPostSimulate` for per-tick logic; `super.EOnSimulate` is the empty stub at
    `enentity.c:201-203`, so calling or skipping it does nothing either way — treat any comment claiming
    it "runs engine/fluids twice" as false. Custom solvers that add `SetEventMask(EntityEvent.SIMULATE)`
    and pump it by hand (aviation) → `dayz-aviation` preflight invariants.
    (hasta 1.29: those `carscript.c` line numbers.) (desde 1.30: `SetEventMask` is `CarScript.c:207-208`; `EOnPostSimulate` is `CarScript.c:807`. Full line-drift table: `references/dayz-1-30-vehicles.md`.)
    
    **Cockpit screen → live engine map (RTTextureWidget bridge) (invariant, added 2026-08-19, SUB_BRZ RT spike):**
    Any hiddenSelections screen can display the ENGINE's map renderer (what LBmaster's GPS uses) instead
    of baked tiles: create the RT via `CreateWidget(RTTextureWidgetTypeID, …)` (children render INTO its
    texture, `enwidgets.c:28` — they never paint to the viewport, so no HUD witness), hang a `MapWidget`
    under it with `CreateWidgets(layout, rt)`, then hook the entity **in this exact order**:
    `SetObjectMaterial(idx, "Mod\data\x.rvmat")` FIRST (no leading backslash), `SetObjectTexture(idx,
    "$rendertarget")` second, `proto native void SetGUIWidget(IEntity, int idx, RTTextureWidget)`
    (`enwidgets.c:637`) LAST. A later SetObjectMaterial rebinds the base texture slot and silently undoes
    the hook — a uniform light-grey screen means `$rendertarget` resolved to the white fallback texture.
    Zero call-sites exist on disk for SetGUIWidget/RTTextureWidget; verified in-game on 1.29 diag.
    MapWidget `GetScale()` defaults to 0.85; at 0.05–0.6 it draws line-work only (contours + roads, no
    urban/forest fill) — pick the zoom with the user's eye gate. If the hooked vehicle leaves the client
    streaming bubble its `GetPosition()` reads `<0,0,0>` and the map recenters to the map corner:
    production code must re-resolve the vehicle every tick, never cache the entity reference. MCP test
    flow that actually seats and starts: `vehicle_get_in_client` (ownership get-in; `vehicle_enter`
    answers seated:1 without sticking) → `engine_set`. Recipe, captures and gotchas:
    `LFVehicleUI_dev\HANDOFF.md` snapshot 2026-08-19.
    **Closed 2026-08-19b — RT unusable for a live GPS:** the RT freezes the FIRST composed frame
    (later `SetScale`/`SetMapPos`/vehicle movement never recompose it), and a MapWidget child only ever
    contributes the vector layers (roads + contours) — the raster fill (land, forest, buildings) never
    reaches the RT, even with a 24 s warm-up between widget creation and the entity hook. The bridge
    recipe above remains valid for STATIC screen content only. Live vehicle GPS goes through baked
    tiles on hiddenSelections instead (next invariant).
    
    **Vehicle screen tiles: terrain satmap, not navigation usermap — and apply server-side (invariant, added 2026-08-19, SUB_BRZ Plan B):**
    Two tile sources share the same `s_XXX_YYY_lco.paa` naming and the same 32×32 grid of 480 m tiles
    on chernarusplus, but look completely different: `dz\gear\navigation\data\usermap\` is the PAPER
    map art (light background, red roads, contour lines — what the ItemMap shows), while
    `dz\worlds\chernarusplus\data\layers\` is the colored aerial satmap (the real "GPS look").
    A GPS screen wants the layers path; world name must be parameterized per terrain. Second half of
    the invariant: `SetObjectTexture`/`SetObjectMaterial` called ONLY client-side is silently reverted
    to the config `hiddenSelectionsTextures[]` whenever the server pushes a visual-state resync —
    observed reproducibly on GEAR CHANGE. **Applying on the dedicated server does NOT fix it
    (refuted in-game 2026-08-19 afternoon): scripted SetObject* state does not replicate
    server→client.** The working pattern is the skin-mod one: re-assert the full scripted visual
    state client-side on EVERY `OnVariablesSynchronized` (no packed==applied early-out; gate only
    the diagnostics). Also force-hide every bank window on the first apply
    (`SetAnimationPhaseNow(name, 1.0)` loop): windows spawn with the config-default texture and
    their initial hide phase is not guaranteed applied. The force-hide only exists at runtime if
    the hide animations actually BAKED — see the skeletonBones invariant below. Satmap-in-bank
    verified in-game 2026-08-19 (SUB_BRZ, 1.29 diag, human driver gate).
    
    **Tile overlap: neighboring satmap layer tiles share 32 px — sample only the tile core (added 2026-09-14, SUB_BRZ s93).** [EXACT] In `dz\worlds\chernarusplus\data\layers\`, column 480 of a tile is column 0 of its east neighbor and row 480 is row 0 of its south neighbor: a per-tile screen map must sample only the core `16/512..496/512` of each tile. Without the crop the measured seam ratio was ~1.9-2.1 and roads jumped ~30 m; with it, 0.90-1.12 (measured offline on the 1.30.164014 Exp satmap tiles). [DESIGN] Name index of `s_AAA_BBB_lco.paa`: AAA is the tile column tx, BBB encodes the row as 31 - tz, and UV (0,0) is the northwest corner.
    
    **Sliding-plate tile window: fix the border row and size the bezel with `wy + COVER + T` (SP-402, added 2026-10-01, SUB_BRZ s95).** [DESIGN] On a sliding-plate tile map with the marker centred, the hidden margin under the window must reach one full tile past the lowest window corner: with a free bottom row it is 1.5 tiles deep (97.5 mm under the centre on a 7x5 grid of 32 mm tiles, window +/-88.63 x +/-52.92 mm, bezel 99 mm). Pinning the bottom row with `minValue`/`offset0` in its `_y` class (same slope), whenever its top edge stays COVER below the lowest projected window corner, cuts the requirement to 88.68 mm and the bezel to 90 mm. Under the window the requirement is always `wy + COVER + T`; with a centred marker nothing below one full tile reduces it. If a smaller bezel is wanted, the lever is tile size (zoom), not the window - say so with a preview.
    
    **Seated-view geometry lives in the View Pilot LOD (invariant, added 2026-08-19, SUB_BRZ v1.2):**
    A car with a View Pilot LOD (resolution 1100) renders THAT copy of the geometry for seated
    occupants, not LOD0. Any geometry/UV/selection patch that must be visible from the driver seat
    (screen UV remaps, marker quads, interior fixes) has to be applied to BOTH res 0 and res 1100 —
    patching only LOD0 looks fine from outside and unchanged from the seat, which reads as "the fix did
    nothing". Seat-only geometry (1PP gap closures) goes in res 1100 alone (SP-469). Also from the same
    session: a tiny 8×8 DXT PAA is invalid (mip chain below 4 px → engine renders the grey fallback);
    for flat-color config slots use a procedural `#(argb,8,8,3)color(r,g,b,1,CO)` instead of a file.
    
    **Animations only bake for selections that are skeletonBones (invariant, added 2026-08-19, SUB_BRZ grey-screen root cause):**
    AddonBuilder silently DROPS every model.cfg animation whose `selection` is not declared in the
    model's `CfgSkeletons` `skeletonBones[]`: the build exits 0, the anim classes stay in the source
    model.cfg, the AnimationSources stay in config.cpp, and `SetAnimationPhaseNow` on their sources
    becomes a silent no-op. Measured differential (SUB_BRZ): `gps_marker` was a bone → its 4 anims
    baked into the ODOL; the 28 `nav_w*` windows were in CfgModels `sections[]` but NOT bones → all
    28 `type=hide` anims dropped, so the untextured bank could never be hidden and rendered as an
    opaque grey panel 0.5–1.85 mm in front of the real screen — a grey that reads as "my texture
    apply fails" while the apply is actually landing fine behind it. Two rules: (1) every animated
    selection — hide anims included — must be listed in skeletonBones[]; being in `sections[]` is
    NOT enough. (2) Verify the bake OFFLINE before deploying: extract the PBO and string-scan the
    binarized .p3d for the animation SOURCE names, which are plaintext in the ODOL (e.g.
    `re.findall(rb'nav_w\d\d_h', odol)` must yield all 28 distinct names). Zero hits = the animation
    does not exist at runtime, whatever model.cfg says. Debugging corollary: the engine grey
    fallback is per-SECTION (`texture_index=-1`); when a screen shows uniform grey, check for
    coplanar untextured geometry IN FRONT of the selection being textured before blaming the
    texture path or the apply call.
    
    **Verify new geometry by its ODOL SECTION, not by strings (invariant, added 2026-08-19, SUB_BRZ marker hunt):**
    Geometry added to a MLOD by tooling can be present in every string table of the binarized
    p3d — selection name, material, texture — and still never render, because whether the
    engine draws a section is decided by fields the strings do not show. Measured on the
    SUB_BRZ dash marker across seven builds:
    
    | Section `special` | Meaning | Renders? |
    |---|---|---|
    | `0x0002C000` | hidden-selection + opaque pass | yes (all working textured sections) |
    | `0x0002C100` | hidden-selection + alpha pass | alpha-dependent, unreliable for tiny quads |
    | `0x0000C000` | no hidden-selection bits | **never** |
    
    Two authoring rules follow. (1) A bone-skinned textured selection must ALSO be listed in
    `CfgModels.sections[]`, or binarize emits `is_sectional=0 / sections=[]` and there is no
    draw call for the bone — being a `skeletonBones` entry is necessary but not sufficient.
    (2) The face texture must be a real `.paa` FILE: a procedural `#(argb,...)` face texture
    on new geometry produces `0x0000C000`. Procedurals are fine in rvmat stages, not as the
    face texture of a section you need drawn.
    
    Gate before every deploy that adds geometry: extract the PBO and parse the ODOL section
    table, asserting `is_sectional==1`, non-empty `sections`, and the `0x20000` bit. The LOD
    address table is found by chaining (`lod_end[i] == lod_start[i-1]`, max end ≈ file size);
    a naive scan for the first plausible table hits a false positive. Reference implementation:
    `<vehicle-import>\work\navscreen_planb\gate_sections.py` (built on an external ODOL→MLOD converter's
    `odol_reader`). String-only scans pass while the geometry is undrawable — that gate cost
    five wasted in-game cycles before it existed.
    
    **PAA files: always encode with ImageToPAA (added 2026-08-19):**
    A hand-rolled PAA whose mip chain runs below 4 px is invalid DXT. The failure mode depends
    on the pass: an 8x8 DXT1 `_co` renders as the engine grey fallback, while a `_ca` with
    2x2/1x1 mips renders as nothing at all on an alpha section. Encode every texture with
    `DayZ Tools\Bin\ImageToPAA\ImageToPAA.exe` from a >=64 px source and let it build the mips.
    
    **Recycle existing faces instead of authoring new ones (invariant, added 2026-08-21, SUB_BRZ marker):**
    On an imported/ripped vehicle model, faces ADDED by tooling can be perfectly formed and still
    never render, while every modification of pre-existing faces shows up reliably. Measured over
    ten builds on SUB_BRZ: a static witness quad — no bone, no animation, bright file texture,
    section `0x0002C000`, winding/flags/normals identical to neighbouring faces that render, and
    nothing occluding it — stayed invisible in-game; a bank-window quad that already existed was
    shrunk, retextured and re-UV'd in the same file and rendered immediately. When a model needs a
    new visible element, look for geometry already in it that can be repurposed (an unused overlay
    quad, a hidden variant, a duplicate panel): you inherit its section, its bone and its proof of
    rendering. Reserve authored geometry for models you build from scratch, and if you must add
    faces to an imported model, prove one renders before building anything on top of it.
    
    **Translation offsets: make the memory axis exactly 1 m and the ambiguity disappears (added 2026-08-21):**
    For `type="translation"` in model.cfg, `offset0`/`offset1` are documented inconsistently: they
    behave as metres in some sources and as a fraction of the axis vector in others, and picking
    wrong is silent — the selection simply moves somewhere off the model, which reads exactly like
    "my geometry does not render". SUB_BRZ lost six build cycles to a marker translated with
    `offset1 = 1.0` against a 169 mm axis: under the metre reading each phase step threw it half a
    metre out of the cabin. Author the memory axis **1.0 m long** in the travel direction and set
    `offset1` to the travel in metres (0.169 for 169 mm of sweep). Both readings then produce the
    same displacement, so the question never has to be answered. When a hidden/animated selection
    is invisible, disconnect its animations and rebuild before blaming geometry or materials — a
    detached control build is one cycle and it splits the two causes cleanly.
    
    **Reading a dark-art dash at night (added 2026-08-19):**
    Instrument-cluster art driven by `dashboardMatOff` (emissive 0) is invisible in an in-game
    night, which reads exactly like "the texture never applied". Pin server time
    (`serverTime="YYYY/M/D/12/0"` in serverDZ.cfg) before judging any interior art gate, and
    give the OFF material a faint emissive so the furniture stays readable after dark.
    
    **Crew / engine state on a non-owner occupant (SP-288, origen LFHeli LF-008):**
    `Car.EngineIsOn()` on a client that is not the Pawn owner returns **false** even when the
    authority has the engine running — no exception, no warning. Measured on the occupant of
    seat 1 with a two-component probe in the same frame (`natEng=false syncEng=true`) and the
    server log showing no stop. Any client subsystem gated on `EngineIsOn()` shuts itself off
    (thermal camera died instantly: LF-008). Remedy: replicate the state
    (`RegisterNetSyncVariableBool` written in `OnEngineStart/Stop`) and have **every** client
    path read `native OR replicated`; an `m_EngineOnAuthority` (OwnerState) snapshot does NOT
    serve passengers. When a client check decides to turn something off, print BOTH sources in
    the same frame.
    
    **Scripted-pilot limitation — `UAInput` has no injection API (SP-276b, added 2026-08-17, LFHeli LF-001):** the dayz-mcp bridge cannot pilot a vehicle whose mod neutralizes the CarController (`SetSteering(0)` every tick) and reads its own UA axes, because `UAInput` does not admit value injection. The working path is a DIAG script in the OWNER that substitutes the UA reading before `WriteToMove`: with the Pawn pattern of active forces the server consumes the axes from the Move (`ConsumeMove → ReadFromMove`), so the script needs no sync and no RPC. Plan v0: `LFHeli_dev/plans/2026-08-17-piloto-scriptado-diag-v0.md`. Status: pending — conditioned on the scripted pilot actually flying (LL-218: do not promote a mechanic without in-game adjudication); SP-276 patch (a) was adjudicated in-game 17-Aug.
    
    **Mirror-vs-rotation invariant (added 2026-08-02, LFHeli HH-60G):** before you declare an
    imported model "mirrored" and reach for the `det=-1` import rule, **check all THREE axes**. A table
    of memory points cannot tell a reflection from a rotation on its own — points have no handedness, so
    only the PATTERN of sign flips does:
    
    **The general test is the SIGNED VOLUME, not counting axes.** Take four LABELLED, non-coplanar
    corresponding points (A,B,C,D) and compute `det[B-A, C-A, D-A]` on each side. An individual point has
    no handedness, but a labelled tetrahedron does:
    
    - signed volume **keeps its sign** -> proper transform (`det=+1`) -> rotation. The model is FINE and
      `det=-1` would ADD a mirror that is not there.
    - signed volume **flips sign** -> improper transform (`det=-1`) -> the reflection is real; apply the
      `det=-1` import rule plus pseudovector (`-S.n`) normals.
    - volume ~= 0 -> your four points are coplanar: pick others, the test says nothing.
    
    Counting inverted axes is only the axis-aligned shortcut of that test, and it is easy to misread:
    
    | axes flipped (axis-aligned case only) | what it is |
    |---|---|
    | ONE inverted, other two intact | reflection |
    | TWO inverted, third intact | rotation of 180 deg about the third axis |
    | all three inverted | point reflection = 180 deg rotation + mirror; decompose before acting |
    
    Real case, and it nearly cost a full re-emission of 16 `.p3d`: one session measured only X, saw the
    sign flipped and concluded "mirrored"; the user authorised de-mirroring. Measuring Z too showed
    **38/38 points with X and Z inverted and Y intact** — a 180 deg yaw (det=+1), which is the pipeline's
    normal convention, not a defect. Control that settled it: the sister airframe that already flies
    in-game has the SAME convention (nose at -Z in the file). Applying `det=-1` would have introduced a
    real reflection where none existed and broken a winding policy already verified in-game. The
    review that settled it did exactly the tetrahedron test: `+6.076679` in file against `+6.076514`
    in runtime, ratio `+0.999973` - same sign, proper transform, argument closed without a game cycle.
    
    **Say what the test does NOT cover.** A proper file->runtime transform does not rule out a mirror
    baked in EARLIER (before the `.p3d`, in the source mesh or the exporter) nor a pre-mirrored UV
    island, which is a texture defect and shows up as backwards lettering with the geometry perfectly
    correct. Those are separate investigations with separate evidence; do not let a clean parity check
    close them.
    
    **Corollary worth more than the rule — prefer RELATIVE measurements.** A comparison made *within one
    file, through one reader* is robust to whatever axis convention your reader uses; an absolute
    file-vs-runtime comparison is not. "The 10 crew proxies carry a frame rotated 180 deg from the 15
    piece proxies" is sound even if you never establish which handedness py3d reports, because both sides
    went through the same lens. When a diagnosis can be phrased as a relative comparison against a
    control inside the same asset, phrase it that way and it survives being wrong about conventions.
    
    **Imported-mesh budget invariant (added 2026-07-29, LFHeli HH-60G):** a **triples/triangle ratio
    measured on a DECIMATED mesh does not extrapolate to an authored one.** `Decimate` collapses edges
    and breaks normal sharing, so it inflates resolved (point, normal, uv) triples per triangle. Measured
    on the same asset: **1.301** on the Decimate-ratio export (0.26-0.6 per group) versus **0.941** on the
    source's own Medium LODs — 28 % apart, and enough to change how many sub-`.p3d` you plan for (2.43
    projected LODs against 1.75 measured). Rule: measure the ratio on the geometry you will SHIP, not on
    whatever export you have lying around. Corollary, and it is the bigger win: **if the source ships an
    authored LOD ladder (GTA `.yft` High/Medium/Low, ripped racing-game, Sollumz-imported `.blend` datablocks), use it
    instead of decimating** — swapping `obj.data` to the `<piece>_medium` datablock is usually better in
    triples AND satisfies the "the agent never decimates visual" clause of a product spec. Re-measure per
    model; never copy a ratio or a resolved ceiling between assets.
    
    **Packaging invariant — a fix that lives in the generator and not in the binary does not exist
    (added 2026-07-29, LFHeli HH-60G):** when the cycle's fix is *adding or changing a PROPERTY* of a
    `.p3d` (`autocenter`, `class`, `sbsource`, `lodnoshadow`), a SHA-based manifest cannot see it — it
    compares the same wrong file on both sides and reports green. Real case: `autocenter=0` was added to
    the hull's visual LOD in both assemblers and the PBO was packed minutes later from `.p3d` built
    before; measured afterwards on the **48 extracted** models, the 8 sub-models carried it and **all 40
    hulls did not**. The post-`ExtractPbo` gate must READ the property, per LOD, on the extracted files.
    Two riders that make the gate survive contact:
    
    - Declare the expectation **in a table with a reason per row, never as "all LODs"**. A healthy pack
      contains **controls that exist in order NOT to change** (an A/B control arm, a byte-frozen
      baseline); a blanket gate rejects them and gets switched off on its first run.
    - The producer must satisfy its own gate. Here "all LODs" would have rejected the canonical
      assembler's own output, because the Memory LOD has no geometry to re-centre and therefore never
      carries `autocenter`. Exclusions are declared, not assumed.
    
    **Deployment-identity invariant — the artifact your gate validates must be byte-linked to the
    artifact the game loads (added 2026-07-29, SUB_BRZ S42):** a native-contract gate that consumes a
    separate ODOL stage says NOTHING about the deployed PBO. `-packonly` copies whatever format sits in
    the compilable tree, and pairing MLOD↔ODOL stage files by homonymous path is not closure. Measured
    case: the gate validated the ODOL stage doors (~2.1 MB each) while the live PBO carried the source
    MLOD (8.7/8.1 MB, byte-identical to the MLOD stage) — nothing in the pipeline linked the two
    artifacts, and the mismatch surfaced only during post-crash analysis. Rule, before ANY deploy of a
    gated artifact: list the PBO, extract or slice every `.p3d` entry by offset, and require (a) the
    expected signature (`ODOL` where the contract is ODOL) and (b) SHA-256 equal to the approved stage —
    fail on wrong signature, missing entry, casefold/slash collision, or hash mismatch. Corollary:
    record the mounted PBO's SHA-256 contemporaneously at every runtime launch; without that seal, no
    post-crash analysis can prove which artifact the process actually loaded.
    
    **Native-crash fingerprint — compare CODE BYTES, not fault addresses (added 2026-07-30, SUB_BRZ
    S42):** the RPT's `Fault address` is not an identifier — ASLR moves the module base, so the same
    code site prints different addresses across runs and reads as "different bugs". What identifies the
    detection site is `Prev. code bytes` + `Fault code bytes` (16 bytes each) plus the engine `Version`.
    Measured case: six `C0000374` crashes split across two addresses (`3523D5F9` and `57B5D5F9`) had
    **byte-identical** code bytes and identical low 16 bits — one code site, one engine build. Reading
    them by address had produced the (false, and formally rejected) claim that the ODOL-era and
    MLOD-era crashes were unrelated; the byte comparison also killed the hypothesis that the artifact
    FORMAT was the discriminating variable, since the crash appeared in both. Report "same/different
    detection signature", never "same/different bug": `C0000374` is detected on free, not where the
    out-of-bounds write happened, so an identical signature still does not prove identical cause.
    
    ## ITERATION BUDGET — check this BEFORE opening a front (added 2026-07-27)
    
    Measured over one month of vehicle work (125 session handoffs: MercedesAMGLF 46, SUB_BRZ 38,
    LFHeli 37): **26 sessions ended in STOP/BLOCKED/FALSIFIED against 24 green**. The three failure
    modes below are process gates, not config facts — and each burned dozens of sessions on a
    different project before being named. They cost more iterations than any invariant below.
    
    1. **Count the cycles spent on one objective, and make the count block.** `DZ-R5` already says
       "after 2 rebuilds without progress, STOP and change strategy"; it does not bite because nobody
       tracks N. Carry `ciclos_en_este_objetivo: N` in the HANDOFF `LIVE-STATE`, and at N>=3 the
       re-entry gate must pick one of three exits before touching code: change strategy, escalate to a
       human, or drop the objective. Evidence: MercedesAMGLF Task 9 reached **v64 across 29 sessions**,
       and what finally closed it was shifting 28 proxy triangles by -0.240 m in X.
    
    2. **Keep the test harness OUT of the product task.** Task 9 mixed "align the Mercedes proxies"
       (product) with "build an automated live verifier" (daemon, broker, lease, preflight, camera
       contract, JPEG SHA) — the harness silently took the bulk of the time and became the project.
       If a product task is blocked TWICE by the harness, stop the product and decide explicitly: scope
       the harness as its own project, or run that one test by hand and move on.
    
    3. **When the judge is the eye, offline analysis gets ONE round.** Green offline gates do not
       predict the live result. A Task 9 build reported `263/263 global + verifier 35/35 hard PASS` plus
       every SHA, and then: `Visual: FAIL of evidence, client hung`. The LFHeli "seis frentes" night
       spent ~25 Codex sessions to produce refutations only (binding intact, re-bake refuted, clock and
       interpolation refuted) and concluded the discriminator was the user's own screenshot. If the
       discriminator is visual or feel, spend one offline round ruling out cheap causes, then go to the
       game — extra offline rounds postpone the verdict, they do not replace it.
    
    Corollary already paid twice: SUB_BRZ needed 30+ sessions to conclude that a decimated rip mesh
    cannot reach game-ready quality (pivot to a human artist), and the LFHeli HH-60G reached the same
    verdict independently weeks later. A visual red that survives N rounds is a signal to escalate to a
    human, not to run round N+1.
    
    > REDIRECT CAMBIO-1: familia B → `../rip-vehicle-import/cookbooks/family-b/radial-puerta-ausente.md`.
    
    ## INV-ALIGN — a faithful import does NOT mean the pieces line up (added 2026-07-27)
    
    **A bit-exact import and correctly-fitting pieces are different properties.** When a model is
    assembled by script, each piece is placed at a **procedurally computed anchor** (centroid, bbox
    centre, a reference point). If that anchor diverges from the real reference, the pieces sit wrong
    while every import check passes. Paid three times (MercedesAMGLF, SUB_BRZ, LFHeli):
    
    - MercedesAMGLF: the 28 wheel proxies sat **-0.240 m in X** from their Memory hubs. Fix: measure
      the offset with py3d and shift only those triangles (max error 0.0 m afterwards).
    - LFHeli OH-1: import fidelity 12/12 at max 3.2e-7 m, and the `SEATS` procedural anchor still
      diverged **583 mm** from the source mesh.
    
    **Gate before declaring an assembly good**: for every proxy, measure its anchor against the real
    reference (the matching Memory point, or the source mesh position) and assert the delta is ~0.
    Import fidelity, digest equality and structural diffs do NOT catch this — they compare the piece
    to itself, not to where it belongs.
    
    **Do NOT "fix" it by adjusting the pipeline transform.** That is the intuitive move and it is a
    false fix: it displaces everything else that was already correct. Correct the individual anchor.
    
    **False friend — `autocenter=0` is not this bug.** It is a real and separate requirement (the
    engine recenters LODs by bbox without it, displacing collision ~20-25 cm from visual; see
    `dayz-model-pipeline` and `dayz-p3d-audit`), but it does not cause or cure anchor divergence.
    MercedesAMGLF spent a full RCA on that hypothesis and closed it as unproven: adding `autocenter=0`
    to all 18 proxy LODs changed nothing geometric, and the protrusions and see-through were already
    present in the earlier baseline. Do not re-run that investigation.
    
    ## TEXTURE OR MESH? — discriminate BEFORE auditing the texture pipeline (added 2026-07-27)
    
    "The textures look wrong" on an imported vehicle is a symptom with at least three unrelated
    causes, and auditing the texture pipeline is the **most expensive** way to find out which. Run
    this ladder first; each rung is minutes, the full audit is sessions.
    
    1. **Audit the UVs, not the textures.** Exact SAT overlap + island count + density
       (`uv-clean-atlas`, `uv_audit.py`). Overlap > 0, atomised islands (2-4 faces each) or wildly
       uneven density means the problem is upstream of any `.rvmat` — no material wiring fixes it.
    
       **No-overlap means an exact `SAT=0` (SP-214, added 2026-08-31).** A Monte-Carlo
       estimate has a measured nonzero floor, so its displayed `0%` is not proof. Route
       vehicle UV repair to `uv-clean-atlas` and use its current chart/SLIM/fold/SAT and
       semantic-packing contract; `PartUV` is not the vehicle default. One anti-fold
       round is a cost target, not a correctness exception: run the same exact SAT
       detector after every round, continue when it is nonzero, and block any mesh that
       does not reach zero. This pack owns the routing and acceptance gate, not a second
       copy of that implementation.
    
    2. **Ask where the mesh came from.** A decimated rip cannot yield game-ready UVs: the cause is
       topology, not the unwrap algorithm or the texture route. The fix is retopo (bake the high-poly
       onto a clean low-poly), and no amount of pipeline correction substitutes for it.
    3. **Render the same material on a known-good control.** If a vanilla-derived control looks right
       with the same route, the pipeline is exonerated and the defect is in this asset.
    4. **Only now** audit `.rvmat` stages, `_co/_nohq/_smdi` paths, `.paa` conversion and
       `hiddenSelections` — see `dayz-texture-pipeline`, and its
       `references/vehicle-materials-and-color-variants.md` for the vehicle-specific shader choice
       (Super where faces carry a `_co`, NormalMapSpecularMap for constant-colour parts, tiled detail
       for overlapping or un-baked UVs).
    
    **Cost rule**: if two rounds of pipeline auditing come back clean, "the texture is applied wrong"
    is REFUTED — change branch instead of running a third. An audit that exonerates the pipeline is
    evidence about where the bug is NOT, and repeating it produces no new information.
    
    Evidence for the rule: on LFHeli OH-1, F6 ran a double audit that exculpated the entire offline
    chain (UVs identical across blend/MLOD/ODOL, old PAA identical to new, `rvmat uvSource=tex`, zero
    `hiddenSelections`, no faces inside the watermark box) — the textures were never the problem. On
    SUB_BRZ the documented root cause was the decimated rip's topology, reached after six rounds and
    a pivot to a human artist. In both cases the ladder above would have branched away on rung 1 or 2.
    
    ### Rip-specific rung 0 — a "paint" material in the source has NO diffuse; never substitute a livery sheet (added 2026-07-27)
    
    Before running the ladder on a **ripped** vehicle (GTA, ripped racing-game, any game rip), census which source
    material each face carries and what the import mapped it to. Game rips colour the body through the
    host game's **paint system**, not through a texture: in GTA those materials are
    `vehicle_generic_smallspecmap*` / "primary" and their diffuse is legitimately **absent or a
    zero-byte file**. An importer that needs *some* texture per face will silently fall back to whatever
    is available — and if what is available is the model's **livery/decal sheet**, the result is signs,
    numbers and camo stamped across the whole fuselage. The symptom reads as "the UVs are wrong" or
    "textures are crossed", and both readings send you down the expensive branch.
    
    - **Discriminator (minutes)**: count faces per `(source material -> texture)` pair. A single texture
      carrying a large share of the model, especially a decal/livery atlas, is the signature. Cross-check
      the source file size on disk: a **0-byte** or missing DDS proves the fallback fired.
    - **Fix direction**: those faces want a **flat colour `_co`** (the host game's paint slot has no DayZ
      equivalent), not the decal sheet and not a re-unwrap. Their UVs were never meant to address a
      texture, so no UV work is involved.
    - **Measured evidence.** LFHeli HH-60G (2026-07-27): `hh60g_sign_1_co.paa` — a 22.4 MB livery sheet —
      carried **23,716 / 35,385 faces (67%)**, of which **8,695 (24.6% of the model)** were pure fallback
      (7,568 of the body's `[PRIMARY]` material plus all four doors), with
      `vehicle_generic_smallspecmap3.dds` confirmed **0 bytes** on disk. The user's independent report was
      "things and colours drawn where they don't belong". The OH-1 line hit the same class separately
      (rotor blades wearing camo and a registration number) after the PAA route had been exonerated.
    - Status honesty: the diagnosis is measured; the flat-colour fix was **not yet confirmed in-game** at
      the time of writing — treat the fix direction as unverified until a cycle closes it.
    
    ## A mesh DRAWN in the wrong place is invisible to every script transform — qualify the probe first (SP-138, added 2026-07-29)
    
    When the user reports that a vehicle **draws** somewhere the entity is not — body on the ground while
    the vehicle climbs, a part left behind, a piece floating — the reflex is to instrument the transforms
    and compare. **That reflex has already been paid for and it returns nothing.**
    
    Measured on LFHeli OH-1 over a 26 m climb: root position, `GetRenderTransform`, both bones, the
    physics body (`dBodyGetWorldTransform`) and the pilot's world position **agreed digit-for-digit**
    while the hull was drawn on the ground. Recorded in that project's own code at
    `LFHeliCore\scripts\4_world\LFHeli\LFHeli_Base.c:3106-3113`. The consequence is the general rule:
    
    > **Any probe derived from the entity root inherits the root's blind spot.** A drawing-position defect
    > lives downstream of every transform Enforce exposes, so no amount of transform logging can see it.
    
    **What does work** — a marker the engine projects from a WORLD point without consulting the entity's
    transform, read off a screenshot: `Debug.DrawSphere` at `GetPosition()` plus a ladder of rungs at
    fixed spacing below it. The rung spacing is what makes the defect **quantitative** (gap in metres =
    rungs × spacing) instead of an eye report. Reference implementation: `LFHeli_Base.c:3114-3128`.
    
    **Probe qualification, mandatory before any number adjudicates:** read the magnitude in the regime
    where the drawing is CORRECT (on the ground) and in the regime where it is WRONG (airborne). If it
    does not change while the eye sees a difference, that magnitude is **blind** — stop using it. A delta
    of exactly `0.000` is a tautology signal, not precision.
    
    **The companion trap, on the inference side.** Over flat ground, "the visual Y is frozen at its
    take-off value" and "the transform is computed from a ground-referenced magnitude" produce **the same
    two signatures**: gap grows with height, gap returns to zero on landing. Arithmetic, confirmed by
    external review. So the flat-ground signature must never be used to EXCLUDE a family of causes — it
    identifies nothing. Discriminate by changing the regime, not by reasoning:
    
    - horizontal traverse at constant world Y over terrain ≥15 m lower — ground-referenced tracks the
      terrain, frozen-Y does not;
    - descend without touching down and climb again — a replication/state split shows hysteresis, a pure
      function of height does not.
    
    - Status honesty: the blind-spot measurement and the flat-ground ambiguity are **verified**; the two
      discriminant manoeuvres above were designed on 2026-07-29 and **had not been run in-game** at the
      time of writing. Treat them as an untested procedure until a cycle closes it.
    
    > REDIRECT CAMBIO-1: familia B → `../rip-vehicle-import/cookbooks/family-b/get-in-ausente.md`.
    
    ## INVARIANTS YOU WILL HIT — preflight checklist (read BEFORE authoring, not after the in-game fail)
    
    These recur on **every** vehicle. They were each won the hard way on one project and then re-derived
    from scratch over dozens of iterations on the next (LFQuad → SUB_BRZ → MercedesAMGLF), because they
    were not promoted to this checklist in time. Check them up front; full detail behind each pointer.
    
    **BEFORE copying a donor — verify WHERE the engine is, not how many seats it has (SP-205, origin 2026-08-08, SUB_WRXSTI T8).**
    `sedan_02.c` looks like the vanilla 4-seat sedan control and is **rear-engine**: it closes `CarBattery`
    and `SparkPlug` behind the **trunk** and the `CarRadiator` behind the **hood**
    (`scripts\4_world\entities\vehicles\inheritedcars\sedan_02.c:142-160`), and gates `CanDisplayCargo`
    on the hood (`sedan_02.c:203`). Porting that to a front-engine car leaves the engine parts behind
    the wrong lid — a bug that only shows in-game. The correct control for front-engine + 4 seats is
    **`civiliansedan.c`**. Rule: before copying a donor, verify WHERE the engine lives, not only how
    many seats it has. Rear seat on that 4-seater (measured vs `civiliansedan_mlod.p3d`; in-game
    confirmation pending WRX live test #1): Memory `pos_cargo1(_dir)`, `pos_cargo2(_dir)`,
    `seat_con_1_2`, `seat_con_2_2`; `GetSeatAnimationType` cases 2/3 → `VEHICLESEAT_PASSENGER_L` /
    `VEHICLESEAT_PASSENGER_R` and **`CrewCanGetThrough` cases 2/3** (`civiliansedan.c:119-131`,
    `:226-256`) — the base returns `false` (`transport.c:727-733`). Distinct `crew_cargo` proxy index
    and seat-box lead: invariants 28 and 30.
    
    **Four-seat completion rider (SP-205, added 2026-08-31).** Rear seats also need
    `seat_cargo1/2` boxes dual-tagged to their exact `componentNN` in Geometry and
    ViewGeometry; `crewcargo1/2` proxies in ViewGeometry and FireGeometry using one
    `crew_cargo` path with distinct instance indexes; `Crew.Cargo1/2` entries with
    `actionSel`, `proxyPos`, `getInPos`, and `getInDir`; and the two passenger cases
    in both `GetSeatAnimationType` and `CrewCanGetThrough`. `CarScript` already maps
    all four `seat_*` selections to `seat_con_1_1`, `seat_con_2_1`,
    `seat_con_1_2`, and `seat_con_2_2` (`carscript.c:2141-2156`), so an override of
    `GetDoorConditionPointFromSelection` that only repeats that table is dead code.
    
    **That base table is the LEFT-HAND-DRIVE mapping (SP-426, added 2026-10-01, WRX STI B-01).**
    [EXACT][CLAIM-GETIN-LHD-BASE-TABLE] On a right-hand-drive car whose memory points are named by
    physical side (driver door = `seat_con_2_1`), the base table pairs every seat with the FAR door:
    each seat is evaluated against the opposite door and `CanReachSeatFromDoors` fails the 1.0 m
    plan-distance test from the correct one, so no get-in appears with names, points and `componentNN`
    all correct. Signature in `query_get_in_condition`: `unreachable` from the correct door and
    `available` from the opposite one. Override it crossing the sides (driver->`seat_con_2_1`,
    codriver->`seat_con_1_1`, cargo1->`seat_con_2_2`, cargo2->`seat_con_1_2`), as vanilla does for its
    own layouts (`OffroadHatchback.c:364-379`, `Van_01.c:269-287`, DayZ 1.30.164014 Exp). On RHD cross
    `GetSeatAnimationType` for the rear seats too (vanilla pairs L with the physical left side,
    `CivilianSedan.c:119-134`); the driver keeps `VEHICLESEAT_DRIVER`, whose get-in animation will look
    LHD-sided - an engine limit, not a mod bug.
    
    **BEFORE any numbered gate below — calibrate the gate itself (SP-132, added 2026-07-29; LFHeli HH-60G + OH-1).**
    Two rules in this file already cover this — §"Calibration + scope (so a gate is trustworthy, not a false
    green)" and §METHOD habit 2 ("never close a phase on a tautological gate"). Both were live and **neither
    fired across four gates in one week**, because both read as being about *car render/winding* gates. They
    are not: they govern every item below, and capacity, parity, telemetry and policy gates alike. The four
    shapes, so they are recognizable on sight:
    - **A threshold copied instead of measured.** `hh60g_assemble.py:29` carried `RESOLVED_LIMIT = 65535`;
      the engine's real resolved-vertex ceiling on that model is **46.133** (42% lower), so the gate passed
      models `binarize` rejects with `Too many vertices`. The cap is model-dependent (×1.4205 on HH-60G,
      ×1.46 on OH-1) — never copy it between models, re-measure with binarize.
    - **A policy that is a mathematical no-op.** `uniform` and `keepdot` were geometrically inert
      (`dot(-cross,-n) == dot(cross,n)`); four variants were built on them before the identity was checked.
    - **A gate that shares the producer's frame.** OH-1 seat parity compared under pure translation — the
      same assumption the sub-model was generated with. `delta 0.000` proved producer and gate agreed, not
      that either was right.
    - **A field read selectively.** `attachment_count=2` was signed PASS because that one field matched
      expectation; the regression sat in the same output 20 min before the user hit it in-game.
    
    **A fifth shape, and it fakes an ABSENCE rather than a green (added 2026-08-19, SUB_WRXSTI): matching
    a LOD by its nominal resolution.** LOD resolutions on disk are not the round numbers the docs use —
    FireGeometry reads `6999999976046592`, not `7e15`. A bucket keyed on `round(7e15)`, or any `==`
    against the nominal, comes back EMPTY, and a gate built on it reports "the selection is absent" about
    something that is present. Measured cost: a cross-check of the config's `componentNames[]` against the
    `.p3d` reported all six damage zones missing, and they were all there in FireGeo — the report reached
    the user as a red before the control against vanilla caught it. Always match with a RELATIVE tolerance
    (`abs(res - nominal) <= nominal * 1e-3`), and note the direction of the failure: this one does not
    show up as a suspicious green you might question, but as a confident red about missing data, which is
    the shape people act on fastest.
    
    **One minute, before trusting any gate:** (a) run it against a KNOWN-BAD case — green on a known-bad
    means the bug is the gate, not the artifact; (b) name in one line the assumption the artifact was BUILT
    with, and confirm the gate does not reuse it. A gate that has only ever been seen green is unmeasured,
    not verified.
    
    - **Gates that stay red stop being sensors (LL-244).** If a known failure is deliberately accepted, pin its exact signature — entities and measurements — and report `KNOWN_RED` separately from `NEW_RED`. Any drift, including an improvement, is new evidence because the pin describes a state, not a ceiling. A gate that returns the same verdict for N runs without making a new red distinguishable has become a constant.
    
    0. **Gate #0 — mesh + UV health BEFORE anything else (SP-052).** For ANY imported model (ripped or not),
       the FIRST step is a mesh+UV audit, because a broken mesh produces false downstream diagnostics (a
       "double wall" that was mirrored UV on a single wall; a "convexity" issue that was open design-boundary
       edges). (a) audit `mesh_health.py` (boundary / non-manifold / dup-verts / dup-faces / winding per
       piece) + `uv_audit.py` (islands / overlap / density / mirrored); (b) sanitize with an ADAPTIVE weld
       (ladder 0.5→0.01 mm; gates: NM must NOT grow + UVs preserved — a blind 0.5 mm weld FUSES
       legitimately-touching pieces and CREATES non-manifold); (c) unwrap BY SECTIONS per
       `AI/20_Knowledge/uv-mapping-dayz.md` §0.5/§5; (d) gates: overlap ~0% for bake, uniform density, atlas
       legible by a human, checker render with SMOOTH shading (flat shading + grid raster read as a false
       "low-poly/blurry"). Tools in `LFQuad_dev\tools\` (mesh_health.py, mesh_sanitize.py bpy-headless;
       uv_audit.py py3d) — portable to any vehicle. → `AI/20_Knowledge/uv-mapping-dayz.md`, winding #10(f).
    
    0b. **Gate #0b — when LOD1/2 come from DECIMATE instead of an artist, audit the modifier's OUTPUT
       (added 2026-08-06, MercedesAMGLF).** No artist is now the normal case, so per-object COLLAPSE +
       planar DISSOLVE is the default LOD path. The modifier leaves **loose vertices** — vertices no
       triangle indexes. Every count-based gate stays green (triangles, resolved, identity, monotonicity,
       frame parity), and the streams ship broken: a per-vertex extract loop driven by `mesh.loops` never
       reaches a loose vertex, so whatever the array was initialised to is what gets written. Initialised
       to `(0,0,0)` → **zero-length normals in NORMAL.bin**. Measured: source had 0 loose in 224.700 verts;
       after decimation, 786 at LOD1 and 4.024 at LOD2 — plus 16 normals of length down to `3e-6` that a
       `> 0` usability test happily passed. Three checks close the whole class:
       - delete loose verts after every decimate (`bmesh.ops.delete(bm, geom=[v for v in bm.verts if not
         v.link_faces], context="VERTS")`). NOT one of the three forbidden calls — it removes only geometry
         no triangle references. Skip the bmesh round-trip entirely when there is nothing to delete, so
         must-keep pieces never pass through it.
       - assert every emitted normal is unit length and normalise on emission. The real cut is **0.5**, not
         0: a unit normal survives float32 within ~1e-7 of 1.0, so anything shorter is a cancelled sum.
         Substitute from NEIGHBOUR normals, never from `poly.normal` — if the import transform negates an
         axis it is a mirror, and the winding-derived normal points the opposite way (see #10 winding).
       - assert every index falls inside its own piece's vertex range. No count gate sees an out-of-range
         index; it explodes at assembly.
    
       **Gate corollary — a drift number that mixes classes is not a measurement.** Split survivors into
       INTACT (incident-triangle set identical to LOD0 → drift MUST be ~0; any drift here is a real defect),
       RECOMPUTED (topology changed → drift is legitimate) and DEGENERATE (counted, never averaged in).
       Mixed together, a p99 of `90.00°` hid a real defect AND hid that untouched geometry was clean at
       ≤0.26°. An exact round angle is a fingerprint, not a result: `90.00` is `acos(0)` = a null vector on
       one side. Match survivors on a ~1e-4 m grid, not 1e-6 — positions are written as float32 and a 1e-6
       grid drops about a third of the untouched vertices at random, which is why the samples looked tiny.
    
    **Position is a candidate bucket, not identity (SP-163, added 2026-08-31).**
    The grid above may find candidates, but a seam can place distinct normal/UV corners
    at the same coordinate. Pair intact data by a carried source vertex/corner and
    primitive ID. If that identity was lost and a bucket is non-unique, return
    `INCONCLUSIVE`, not PASS or FAIL. Run source against itself through the same matcher
    and require approximately zero residual before trusting any product comparison.
    
    1. **"Get in" radial needs a real SCRIPT CLASS, not a bare config class.** A `class <Mod>: CarScript`
       with no `.c` runs as bare `CarScript`, which inherits `Transport.CrewCanGetThrough()` → **`false`**
       (`scripts/3_game/vehicles/transport.c:727`) → the get-in action is filtered out and **never appears,
       no bone/proxy/componentNN fix helps**. Author `<Mod>_Base extends CarScript` overriding
       `CrewCanGetThrough` + `GetAnimInstance` + `GetSeatAnimationType` (+ `CfgMods` worldScriptModule).
       Verified in-game LFQuad D34 + MercedesAMGLF + SUB_BRZ. → `vehicle-structural-parity.md` "Crew get-in".
    2. **The script module must actually LOAD or the class never binds — silent.** `CfgMods` `files[]` with
       back-slashes / a `.p3d` path that resolves to `*.p3d.p3d` → module not loaded → no get-in, no error.
       → `SKILL.md` §"Binding del script", `rip-import.md`.
    3. **REDIRECT CAMBIO-1 (familia B):** `../rip-vehicle-import/cookbooks/family-b/wheelpresent-0.md`.
    4. **Seats and wheel hubs must be `componentNN`-tagged (dual-tag) — AND each seat component must be raycast-collidable.** The engine enumerates collision
       components only by `componentNN`; a seat/hub that is a standalone island (0% componentNN overlap) is
       invisible → spawn blocker / no seat. Tag the SAME faces with a `componentNN` too. **Then the ViewGeo seat cube
       must be INWARD-wound with every point flag = `0x0000003F`:** a py3d box left outward + flags 0 passes every
       offline shape/winding/dual-tag check yet `RaycastRV(ObjIntersectView)` does NOT hit it → the get-in cursor
       falls through to component0 → the driver "works" by fallback but the CODRIVER never resolves. The decisive,
       in-game-confirmed copilot blocker on BOTH SUB_BRZ (s9) and MercedesAMGLF (s12, headless `hit=1 comp=6 crewIdx=1`). The s9 patch changed winding and flags together and never isolated the flag as cause; the safe rule is to copy the sealed vanilla control's convention, never py3d's default. → "componentNN DUAL-TAG" + "CRITICAL EXTENSION 2026-06-28" (`vehicle-structural-parity.md`).
       positive control (LFQuad/Croco), never trust py3d's default. → "componentNN DUAL-TAG" + "CRITICAL EXTENSION 2026-06-28" (`vehicle-structural-parity.md`).
       **Resolve the seat's `componentNN`; never hardcode its number (SP-079, added
       2026-08-31).** Assembly order moves component numbers. For every `seat_*`
       selection, compare both point and face membership against each `componentNN` in
       the same LOD and require exactly one exact 1:1 match. Zero or multiple matches is
       a hard failure. Re-run the equality after any transform or transfer so the
       seat/component dual tag cannot be lost silently.
    
    5. **Crew/wheel proxies must exist in BOTH ViewGeometry AND FireGeometry**; the proxy triangle uses the
       engine identity frame (`R=((-1,0,0),(0,0,1),(0,1,0))`, model-space), NOT py3d `rotation=None`. → parity + rip-import.
    
    5b. **`wheelHub` names a GEOMETRY/MEMORY selection, not the visual `wheel_X_Y` — and the visual
       `wheel_X_Y` IS the proxy's own triangle under a second name (added 2026-08-07, MercedesAMGLF).**
       `class Wheels { class Left { wheelHub="wheel_1_1_damper_land"; } }` resolves to a selection that
       lives in Geometry `1e13` (a ~0.20 m hub box) and Memory `1e15` (1 point) — vanilla
       `DZ/vehicles/wheeled/config.cpp:297`. It is NOT the `wheel_X_Y` selection on the visual LOD.
       Measured on two independent cars: that visual companion has the SAME centroid and the SAME bbox
       as its `proxy:<path>.NNN` selection, i.e. it is one triangle carrying two names (vanilla shows it
       as 1 face too). Consequences, each of which cost a session:
       - **A gate of the form "proxy anchor -> companion centroid == 0" measures nothing about the wheel.**
         It computes `|anchor - centroid|` of a single triangle, which is a pure function of that
         triangle's SIZE. Measured on the same model: 0.131757 m for a 0.339x0.204 triangle vs
         0.000745 m for py3d's canonical `scale=0.001`. Against the real hub both artifacts read
         **0.000000 m**. Point the gate at `wheel_X_Y_damper_land`.
       - **Proxy triangle size bakes NO scale.** The frame is derived from UNIT vectors
         (`py3d/__init__.py:174-179`, port of `proxy_frame.derive_frame`), and vanilla's binarised proxy
         matrix is orthonormal (`|aside|=|up|=|dir|=1.000000` on `civiliansedan.p3d` ODOL v54). A 1 mm
         canonical triangle and a 0.34 m one place the same wheel; SUB_BRZ shipped on `scale=0.001`.
         Do not "fix" a small triangle.
       - **Vanilla reference, if you need one:** `civiliansedan` puts the hub **0.106334-0.107333 m
         INBOARD** of the wheel proxy, the whole delta in X. Hub coincident with the proxy (0 m) is
         tighter than vanilla, not a defect.
       - **Before matching a magnitude to another open debt, decompose it by AXIS.** 0.1318 m was read
         as "the wheel drawn 13 cm off its simulated position" because it resembled a 0.136 m track
         debt; the 0.1318 lives entirely in Y/Z (`dX = 0.000000` exactly on all four wheels) and the
         track debt lives entirely in X. Orthogonal — the resemblance was numerology.
       - **Read a gate's PRODUCT verdict from a run WITHOUT its negative fixtures**, then repeat it on
         the known-good artifact as a control. Run with `--negative-fixture` and the fixture's
         deliberate reds (yaw-180 families, stripped companions) interleave with the product rows and
         read as product failures.
    
    5c. **Owner-only driving hitch: audit hub-ground clearance and remove `LandContact`
        before blaming scripts (SP-355, added 2026-08-31).** The driver owns the physics
        solve while a passenger interpolates, so smooth passenger motion does not clear
        collision geometry. Working DayZ car controls use the tyre's PhysX sphere plus a
        small `wheel_*_damper_land` Geometry box (about 0.20×0.22×0.22 m) and no
        `LandContact` LOD. Require `hub_y - mounted_tyre_radius` to exceed both
        `Suspension.travelMaxUp` and `tra

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.