fix: stop shrinking CAD models rotated off-axis

[addibble]

Problem

A CAD model that declares a size matching its own mesh is shrunk whenever its component sits at a rotation that is not a multiple of 90°.

This 6×6mm tactile switch (9.5mm lead span) is the same part four times, at 0°, 30°, 45° and 90°. Only the right-angle ones meet their pads:

rotation	rendered scale
0°	1.00
15°	0.56
30°	0.42
45°	0.39
90°, 180°	1.00

Cause

getObjectBoundsRelativeToParent measured each mesh by transforming its bounding box into world space and then back into the parent's frame, as two separate calls:

transformedBounds.applyMatrix4(node.matrixWorld)
transformedBounds.applyMatrix4(parentInverseMatrix)

Box3.applyMatrix4 re-derives an axis-aligned box around the eight transformed corners, so it is not invertible. Applying a rotation and then its inverse doesn't cancel — each call grows the box:

9.50 × 6.05  ──rotate +30°──►  11.25 × 9.99  ──rotate −30°──►  14.74 × 14.28

getCadModelFitScale then divides the declared size by that inflated measurement and shrinks the model to 0.42 to make it "fit".

Right angles are unaffected because an axis-aligned box is rotation-invariant there.

Fix

Compose the two matrices and apply them once, so the ancestor transform genuinely cancels:

const localMatrix = parentInverseMatrix.clone().multiply(node.matrixWorld)
transformedBounds.applyMatrix4(localMatrix)

Rotations at or below the measured object are still included — the existing fit includes loader transform group rotation in measured bounds test covers that and still passes.

Why this went unnoticed

Two things hid it:

• size is optional, and the fit is skipped without it (if (!targetSize) return [1,1,1]). Parts that don't declare one were never affected.
• The headless pipeline is correct. circuit-json-to-gltf applies component rotation as a scene-graph property after fitMeshToCadBounds, so its fit always measures an unrotated mesh. I confirmed it returns 9.500 × 15.000 × 6.050 at 0°, 30° and 45°. Snapshot tests, tsci build and poppygl renders are all correct — so no existing test could have caught this. It is visible only in the browser viewer.

Tests

• tests/cad-model-fit.test.ts — asserts scale 1.0 at 0/15/30/45/90/180°. Verified it fails without the fix.
• stories/Bugs/RotatedModelFitScale.stories.tsx — visual regression. Pads use rotated_rect with ccw_rotation and the silkscreen outline is rotated point-by-point, so both turn exactly with the part and a shrunken model visibly pulls away from them. Each instance declares size deliberately, since the fit is skipped without it.

Full suite matches main exactly: the 4 pre-existing failures (faux-board z defaults, Atari SVG outline) are unchanged, and the fit tests go 4 → 5 passing.

No new Circuit JSON properties are used; everything is in the pinned circuit-json@0.0.446.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
3d-viewer	Error		Aug 5, 2026 1:34am

[tscircuitbot]

---

Thank you for your contribution! 🎉

PR Rating: ⭐⭐
Impact: Minor

Track your contributions and see the leaderboard at: tscircuit Contribution Tracker

---