[boazeggs]

Point-colour texturing previously interpolated the reconstructed mesh's stored vertex RGB, and photo visibility only considered the reconstructed faces. This could retain stale colours and project a photograph through foreground measurements omitted from the mesh.

Resample original cloud RGB with a bounded spatial index, distance weights, local surface normals and robust colour rejection. Respect the layer's colour/deletion snapshot. Unsupported areas remain explicitly unobserved instead of inheriting coarse vertex colour. Add bounded source-cloud depth maps to photograph visibility, including foreground outside the reconstructed ROI. API revision 18 reports measured versus vertex fallback, source sample count and sampling limits; the native Properties view shows the sampling limitation.

Closes #13. Refs #32.

This PR builds on #62 at 5423c2859ead73f4f65030ec1a374c0b3a1fdc07, which includes #53 and the earlier integration. The incremental texture commit is 0962214392f806bdf4c95bcafaecce081bfd2276. Merge #53 and #62 first for an incremental review; their review heads remain unchanged. Existing PR heads remain ancestors, so use normal merge commits when approved.

Validation:

• Entire Rust workspace: 1,480 passed, 0 failed, 12 ignored; workspace/all-target and minimal-core Clippy pass with warnings denied. Rustfmt, Python syntax, locale JSON and diff checks pass.
• Measured colour ramp stays within three RGB levels. Tests reject an isolated nearest-colour outlier, wrong surface directions, a nearby parallel wall and source-cloud photo occlusion outside the mesh. Atlas pixels repeat deterministically; edited/deleted colours and source identity are tested.
• Actual native viewer: 903 source points produce 1,800 measured RGB texels and 756 unobserved texels, with no stale red vertex fallback. Repeated PNG bytes match; toggling the measured atlas changes 47,957 viewport pixels. The original fixture remains unchanged.
• Real Riegl E57: 1,095,702 points and five station images; 43,913 photo texels, no failed decoding, safe cancellation and unchanged original SHA-256.
• Real BLK2GO E57: 199,186,802 source points, 262,144 bounded samples, 23.63 s bake and 660,152 KiB peak native RSS for reconstruction/baking together. Progress advances, cancellation leaves no files, and the source SHA-256 stays unchanged. OBJ/MTL/PNG reopens and renders in the independent trimesh/pyglet OpenGL viewer, as does the generated colour fixture.

Repeatable probes, hardware/settings and recorded results include the exact limits. Sampling is capped at 262,144 points; without an octree it uses a bounded resident preview. Photo depth maps are at most 512×512, cached for eight images, so empty cells make no visibility claim. The deliberately coarse 5,000-vertex large fixture has 9,432 open edges and many unobserved texels: this validates bounded processing and explicit missing coverage, not reconstruction quality or full SLAM trajectory-photo coverage. Timings are shared-workstation development runs, not controlled performance comparisons. Vendor scans/images are not redistributed.