Interior-texel applicability regression (#4381)

The real captured-boulder UI exposed an applicable transfer whose exported atlas contained only the target's old blue appearance. Its positive coverage came entirely from triangle centroids; none of the interior raster pixel centers had an accepted scan observation. Centroid coverage alone is therefore insufficient for an applicable texture mutation.

comparison.json records the exact captured native request, effective IFC and RGBA input hashes, runtime hashes, native results and independent Pillow PNG inspection. Inputs were captured before Apply from the normal transfer UI. The source is the same CC0 Poly Haven boulder qualified in the original transfer evidence; all 66,122 source triangles were retained. The selected IFC target is a 350-triangle captured region prepainted blue. This is a derived scan control, not independent scan/BIM registration-accuracy evidence.

The sparse control uses 64 texels/metre and 0.001 m ambiguity distance. Before the fix, 324 of 890 observations were accepted and the plan was applicable. Independent PNG decoding finds exactly two RGBA colors: the old blue (23,70,223,255) and transparent padding. After the fix, the same geometric estimate remains visible, but the report distinguishes 324 observed centroids from zero observed interior texels out of 540. There is no applicable plan or PNG asset.

The dense control uses 512 texels/metre and 0.00001 m ambiguity distance. Both versions produce the same PNG bytes. The new report counts 6,117 observed interior texels out of 7,963, separately from 350 observed centroids. Independent decoding finds 8,151 colors, including scan albedo and preserved old blue pixels. Changing both controls establishes the positive case; it does not isolate density as the sole cause of improved observations.

Verification and scope

The native regression constructs a small source patch around a target triangle's centroid, proves that its sparse shared atlas contains only the old blue pixels, and refuses Apply. A dense version proves that red source pixels and preserved blue pixels both reach the PNG. The real WASM contract repeats centroid-only refusal and the dense positive case. Existing unknown-area estimates and sample counts remain unchanged; additive counters expose how they were obtained.

To repeat the public-model control, capture the transfer worker's frozen request, effective IFC bytes and RGBA buffer before Apply, run the same payload through IfcAPI.planMeshTransfer with the two density/ambiguity pairs above, unpack IFPA, and decode the resulting PNG independently. Exact local acceptance inputs are retained under /tmp/transfer-texel-real-input; they are not shipped as fixtures. The comparison's input hashes prevent substituting a different request or model. The canonical load path, geometry, nearest-surface predicates, work cap and texture sampling are unchanged by this fix.

Ordinary-load performance

native-load.json contains exact-base/branch/base/branch probes, five iterations each, on the qualified AC20 fixture. Base source is 22f3f0d0d; branch source is fd88626ad. Binaries were built separately, copied immutably, and measured while all collaborating agents held heavy work. The first pair's parse/geometry/total changes were +1/+1/+1 ms; the second pair's changes were 0/-1/0 ms. These mixed phase differences provide no consistently directed change. All runs retain 285 meshes, 35,940 vertices and 19,456 triangles. Separate untimed-for-this-comparison checks retain ordered mesh FNV 25ac885b6ff4ad00 on both builds. That fingerprint covers mesh payloads, not textures; dense transfer PNG identity is independently recorded in comparison.json. No browser worker speedup or expanded transfer capacity is claimed.