FAIL-MATRIX-WORLD-STALE

Four occurrences in one session · ~2 hours

Setting a transform and then reading matrix_world returns the old matrix. Every calculation built on that read is computed from a fiction. It produced a 109-metre leaf spring on Day 2 morning, a jockey 1.36 m from where it belonged, an awning whose parts sat at double height, and six wheels collapsed onto the world origin.

The insidious part: it was documented the same morning and used again four times that evening, because each new problem looked different from the last.

FAIL-OVERWROTE-PARENT-INVERSE

Destroyed data · ~1 hour to reconstruct

Re-parenting six wheel assemblies, the child matrix_parent_inverse values were reset to identity on the assumption they were redundant. They were not: on those wheels the parent inverse was the placement data — position, scale and the left/right mirror. All six wheels collapsed to the origin, the trailer wheels silently shrank from 778 mm to 741 mm, and one side ended up inside-out showing the hub face.

FAIL-REPARENT-ON-DISPLACED-FRAME

121 objects, 26 m out

Objects were adopted into an animation rig while the timeline sat on frame 1 — where that rig is 26 m off-camera. Preserving their world positions baked a +26 m offset into every local transform. The car's wheels, the coupler hierarchies and the towball were left standing on the beach while the car drove away.

FAIL-PARENT-CHILD-COLLECTION-SPLIT

A trailer with no wheels, shipped

The wheel assemblies were empties filed in 03_RunningGear, but all 52 of their mesh children still sat in Reference_CC-NC from the day they were copied off the tow car. The licence-exclusion checkbox stripped the children and left the empty parents. The exported model reached the browser wheel-less — and the export was correct; the filing was wrong.

FAIL-CONSTRUCTION-GEOMETRY-VISIBLE

A mystery stipple line, twice investigated

Boolean cutters for the door openings were hidden from renders but left visible in the viewport, sitting exactly coplanar with the body skin. The result is a dashed white line along the panel — classic z-fighting, and easily mistaken for a modelling fault. A studio backdrop plane left over from a product render did the same thing to the sand.

FAIL-ORIENTATION-UNVERIFIED

Car hitched nose-first; doors swinging into the body

Two orientation errors survived because they were asserted rather than measured. The tow car was hitched by its front bumper. The rear doors were animated to swing into the trailer, because the rotation sign was assumed instead of tested. Both were caught by eye, from a viewport, by the operator — not by the numbers.

FAIL-PROCEDURAL-DOESNT-EXPORT

Flat grey where the detail was

The chequer plate, the solar cells and the awning canvas were all procedural node graphs. glTF has no procedural nodes, so all three exported as plain slabs. This is not a bug; it is what a delivery format is.

FAIL-EXPORT-ALL-SCENES

A stowaway working copy, 100 KB

Blender's glTF exporter writes every scene in the file unless use_active_scene=True. A working copy of the jockey, living in a separate part scene, silently rode along in the published model. Debug collections do the same unless explicitly excluded.

FAIL-SAVED-TO-BACKUP-PATH

Two hours of work invisible in the main file

A recovery step opened trailer_day2.blend1 — the backup file — and Blender then kept saving to that path, creating .blend11 as its own backup. For two hours the canonical trailer_day2.blend was frozen at an old state. Anyone opening it would have found the evening's work missing.

FAIL-COMPOUND-INSTEAD-OF-REVERT

The whole two hours

When the first wheel fix produced a wrong result, the response was another fix, computed from the same broken state. Then another. Four attempts, each inheriting the previous one's damage. The correct move after the first failure was to stop, restore, and change method — not to push forward.

FAIL-NO-SNAPSHOT-BEFORE-BIG-CHANGE

No clean restore point existed

Version discipline says snapshot before every significant change. It was skipped before the animation work began. When a rollback was finally requested, no file existed containing the evening's model work without the damage — the only choice was to strip the animation surgically rather than revert.

FAIL-TOO-MANY-CHANGES-PER-STEP

Impossible to isolate the fault

One script built the drive path, wheel rotation, trailer articulation, body pitch, the camera and the terrain mask together. When the result was wrong, nothing could be attributed. Six variables, one observation.

FAIL-SILENT-RUN-30MIN

Trust

A long sequence ran without progress reporting. Thirty minutes of silence during work the operator could not see. Separately: several results were reported as complete before being looked at — the geometry claim asserted from a return value rather than a render.

FAIL-VERIFY-BY-LIT-RENDER

Day 1, still relevant

Geometry checked in a lit render hides in shadow and specular. A blanket normals recalculation across 167 meshes fixed 42 and silently inverted 52 non-manifold wheel shells, where "outside" is undefined and the heuristic guesses.

FAIL-HIDE-RESTORE-NOT-IN-FINALLY

838 objects · shipped models missing the body

A probe render hid everything except its subject, then errored on BLENDER_EEVEE_NEXT — an engine name that no longer exists in Blender 5.1 — and never reached its restore line. Every later script then captured that broken state as its "previous" state and faithfully restored it. 838 objects were left hide_render=True, body_shell among them, and the exporter drops hidden objects. Models published in that window shipped without the trailer body.

FAIL-SOLID-BOX-WHERE-SHELL-NEEDED

Three builds · nothing rendered inside

cfx5_body was a solid cube, entombing its own liner and wire basket. cfx5_gasket was a solid slab across the aperture, sealing the tub. And body_shell is the same class of object — the 12-vertex solid already logged in the import gate. Cutting the felt for the prow aperture only exposed the inside of that box, which reads as a flat dark panel and looks like a texture fault.

FAIL-STALE-MATRIX-WORLD (×4)

Four occurrences in one day

The bridge's own documented trap, hit four separate times. Rear-door MDF infills landed at rig-local coordinates in world space, hanging under the chassis. A fridge lid sat 400 mm high because the rig's matrix_world was still identity when its children captured it. A re-imported BBQ arrived at 308 metres. The prow parenting went the same way.

FAIL-BOOLEAN-BEFORE-SOLIDIFY

Body collapsed to 8 vertices

An aperture cutter was added to body_shell and pushed to the top of its modifier stack, ahead of the BEVEL and SOLIDIFY. The evaluated mesh dropped to 8 vertices — the whole body reverted to a bare cube. The seven existing cutters all sit after the solidify; that ordering is how the shell is generated, not housekeeping.

FAIL-REGEX-KNEW-HALF-THE-CONVENTION (×3)

Nine lamps dark · twelve panels uncoloured · a BBQ invisible

Three versions of one mistake, all in the viewer rather than the model. The night regex matched led_* and lens_* but not the nine emis_* lamps. The felt regex was /^felt_\d/, silently skipping felt_partition, felt_riser and felt_reardoor. And a BBQ option was defined but never added to the shown list, leaving 23 parts in the file with no control to switch them on. In all three cases the geometry shipped correctly and the consumer was never told about it.

FAIL-BLOCKER-ASSERTED-FROM-HIDDEN-GEOMETRY

A design decision argued from a phantom

A fridge extraction route was declared impossible partly because a jerry cage occupied the path. Those objects were eye-hidden — not in the exported model, not part of the design. A clash test that includes hidden geometry answers a question nobody asked.

FAIL-REASONED-PAST-THE-INSTRUCTION

Two rebuilds

Told a slide was “lengthwise, not sideways”, I analysed sideways anyway. Told a swing-out unit was “exactly the same as the prow door”, I moved its hinge to the opposite end and spent two rounds rediscovering that the door's own pivot — 44 mm proud of the face — was the answer. Both constraints had already been supplied.

FAIL-INSULATION-OUTSIDE-THE-WALL

Four screenshots missed it; one raycast found it

A six-direction ray sweep over a grid of interior points, flagging any first hit on body_shell, returned exactly one exposure: interior_insul_front built at y 1399–1424 against a shell inner face at 1398. The nose insulation was outside the wall, and the felt added on top of it inherited the error.