fix: render Tekla rebar IfcSweptDiskSolid bars correctly (#1348, #1350) (#1351)

* fix(geometry): sample trimmed-circle directrix arcs in 3D (#1348)

IfcSweptDiskSolid rebar (IfcReinforcingBar) whose directrix is an
IfcCompositeCurve with arc bends rendered mangled — the L-bar grew a
spurious hook, the U-bar twisted into a crumpled mess.

Root cause: a directrix segment that is an IfcTrimmedCurve over an
IfcCircle/IfcEllipse was sampled through the 2D conic path and lifted
with z=0, dropping the arc's out-of-plane component. The bend arcs in
these files live in the XZ plane (the circle's IfcAxis2Placement3D has
a non-Z axis), so the flattened arc landed in the wrong plane and broke
the directrix that the disk is swept along.

Fix: route trimmed circle/ellipse directrix segments through a new
process_trimmed_conic_3d that samples the arc against the conic's full
3D placement (center + X/Y axes), honouring parameter and cartesian
trim bounds and matching the 2D sampler's segment-density budget. The
3D placement reader is shared with the existing full-circle sampler;
the planar (IfcAxis2Placement2D) branch now also honours RefDirection.
Both the trimmed and non-trimmed composite-curve directrix paths route
through the fixed sampler.

Adds neutral, self-contained L-bar and U-bar fixtures (faithful copies
of the reported geometry chains) plus regression tests asserting the
directrix stays in its authored plane and the swept tube stays thin.

* test(geometry): add crank-bar (shape 26) regression for #1350

Issue #1350 (crank bar renders nothing when viewed alone) shares the
#1348 root cause: a trimmed-circle directrix arc flattened to z=0. Here
the two bends are tiny (~4.7°), so the pre-fix mangling was subtle
(directrix Y deflection ~10.7 mm) but the resulting self-intersecting
tube degenerated enough that the viewer dropped it entirely.

Adds a neutral crank-bar fixture (faithful copy of the reported
geometry: line, tiny arc, offset run, tiny arc, line) plus a
directrix-planarity test (catches the regression: max |y| 10.7 mm pre-fix
vs ~0 fixed) and a renders-non-empty test. With the directrix sampled in
3D the crank produces a clean 384-triangle swept tube.

* fix(viewer): keep small high-aspect elements compact-framed (#1350)

A single IfcReinforcingBar viewed alone loaded correct geometry but
rendered as nothing. The camera-fit policy classifies any bounding box
with aspect > 50:1 as "linear infrastructure" and frames it from inside
the bbox looking down the longest axis — the right pose for a 900 m
railway alignment, but for a 4.86 m rebar (aspect ~130:1) it points the
camera end-on down the bar so it projects to a sub-pixel smear.

Gate the linear policy on an absolute size floor: the longest axis must
be at least 100 m. Below that the compact SE-isometric pose frames the
whole element (as it already did for the lower-aspect L-bar). The floor
is overridable for tests. Railway/road alignments (hundreds of metres)
are unaffected; the smallest existing linear-branch test is 100 m.

Adds changesets: the swept-disk arc 3D fix ships in @ifc-lite/wasm and
this camera-fit fix in @ifc-lite/renderer. Together with the geometry
fix this closes #1350.

Author: louistrue

.changeset/camera-fit-small-linear-element.md

@@ -0,0 +1,5 @@
+---
+"@ifc-lite/renderer": patch
+---
+
+Keep small high-aspect elements on the compact camera-fit pose. The linear-infrastructure fit policy (camera positioned inside the bbox looking down the longest axis) is meant for railway / road alignments hundreds of metres long, but it triggered on any high-aspect bounding box regardless of absolute size. A single reinforcing bar viewed alone (e.g. a 4.86 m bar, aspect ~130:1) got framed end-on from inside its own bounding box and rendered as nothing (issue #1350). The linear policy now requires the longest axis to be at least 100 m; below that the compact SE-isometric pose frames the whole element. Fixes the rendering half of #1350.

.changeset/swept-disk-trimmed-arc-3d.md

@@ -0,0 +1,5 @@
+---
+"@ifc-lite/wasm": patch
+---
+
+Sample `IfcSweptDiskSolid` directrix arcs in full 3D. A directrix segment that is an `IfcTrimmedCurve` over an `IfcCircle`/`IfcEllipse` was sampled through the 2D conic path and lifted with `z = 0`, dropping the arc's out-of-plane component. Rebar bend arcs (Tekla `IfcReinforcingBar` bodies) live in the XZ plane, so the flattened arc landed in the wrong plane and twisted the swept tube — L-bars grew a spurious hook and U-bars crumpled (issue #1348). The arc is now sampled against the conic's real 3D placement (centre + X/Y axes), honouring parameter and cartesian trim bounds. Fixes #1348 and the geometry half of #1350.

packages/renderer/src/camera-fit-policy.test.ts

@@ -56,6 +56,38 @@ describe('pickFitPolicy', () => {
       assert.strictEqual(policy.kind, 'compact');
       assert.strictEqual(policy.aspect, 10);
     });
+
+    it('keeps a small high-aspect element compact (single rebar, issue #1350)', () => {
+      // A 4.86 m × 0.084 m × 0.037 m reinforcing bar viewed alone — aspect
+      // ~131:1, well past the linear threshold, but only metres long. The
+      // linear "look down the axis from inside the bbox" pose framed it
+      // end-on and it rendered as nothing. The absolute size floor must keep
+      // it compact so the SE-isometric pose shows the whole bar.
+      const policy = pickFitPolicy(
+        bounds(2005.06, 5961.1, 25.5, 2005.097, 5961.184, 30.36),
+        { fovY: FOV_45 },
+      );
+      assert.strictEqual(policy.kind, 'compact');
+      assert.ok(policy.aspect > 100, `aspect ${policy.aspect} should exceed 100`);
+      // distance = longest * 2; longest ≈ 4.86, so ≈ 9.72 — a normal
+      // see-the-whole-bar framing, not an inside-the-bbox linear pose.
+      assert.ok(
+        policy.distance > 9 && policy.distance < 11,
+        `distance ${policy.distance} should be ~2x the 4.86 m length`,
+      );
+    });
+
+    it('honours the linearMinLongest floor override', () => {
+      // 80 m × 1 m × 1 m — aspect 80, above the aspect threshold but below
+      // the 100 m default size floor → compact. Lowering the floor to 50 m
+      // lets it cross into linear.
+      const b = bounds(0, 0, 0, 80, 1, 1);
+      assert.strictEqual(pickFitPolicy(b, { fovY: FOV_45 }).kind, 'compact');
+      assert.strictEqual(
+        pickFitPolicy(b, { fovY: FOV_45, linearMinLongest: 50 }).kind,
+        'linear',
+      );
+    });
   });
 
   describe('linear branch', () => {

packages/renderer/src/camera-fit-policy.ts

@@ -30,6 +30,21 @@ import type { Vec3 } from './raycaster.js';
 /** Aspect ratio above which a model is treated as linear infrastructure. */
 const LINEAR_ASPECT_THRESHOLD = 50;
 
+/**
+ * Minimum longest-axis extent (world units, i.e. metres) for the linear
+ * policy to apply. The linear "look down the longest axis from inside the
+ * bbox" pose only makes sense for genuine infrastructure (railway / road
+ * alignments are hundreds of metres). A small but high-aspect element — a
+ * single 4.86 m reinforcing bar is aspect ~130:1 — would otherwise get the
+ * linear pose, which places the camera *inside* its bounding box looking
+ * end-on, so the bar projects to a sub-pixel smear and reads as "nothing
+ * rendered" (issue #1350). Below this floor the compact SE-isometric pose
+ * frames the whole element and keeps it visible. Picked so the longest
+ * building elements (rebar, steel members, long beams ≲ tens of metres)
+ * stay compact while alignments (≥ hundreds of metres) stay linear.
+ */
+const LINEAR_MIN_LONGEST = 100;
+
 /**
  * Target on-screen projection for the smallest non-degenerate dim, in
  * pixels. The linear-policy distance is chosen so the shortest meaningful
@@ -78,6 +93,12 @@ export interface PickFitPolicyOptions {
    * at the default; exposed for tests that pin the threshold behaviour.
    */
   linearAspectThreshold?: number;
+  /**
+   * Override the minimum longest-axis extent (world units) at/above which the
+   * linear policy is allowed to apply. Production should leave this at the
+   * default; exposed for tests that pin the size-floor behaviour.
+   */
+  linearMinLongest?: number;
 }
 
 /**
@@ -106,8 +127,14 @@ export function pickFitPolicy(
   const shortest = Math.min(sizeX, sizeY, sizeZ);
   const aspect = longest / Math.max(shortest, longest * 1e-6);
   const threshold = options.linearAspectThreshold ?? LINEAR_ASPECT_THRESHOLD;
-
-  if (aspect <= threshold) {
+  const minLongest = options.linearMinLongest ?? LINEAR_MIN_LONGEST;
+
+  // Compact unless the bbox is BOTH high-aspect AND large enough to be real
+  // infrastructure. The size floor stops a small high-aspect element (a
+  // single rebar / steel member) from getting the "look down the axis from
+  // inside the bbox" linear pose, which renders it end-on and invisible
+  // (issue #1350).
+  if (aspect <= threshold || longest < minLongest) {
     // Compact: reproduce the legacy SE isometric pose 1:1 so building
     // models frame exactly as before.
     const distance = longest * 2.0;