Need mesh↔shape deviation (Hausdorff) tool + ability to measure invalid/loose-face shapes

[gsdali]

Context

Using OCCTMCP to verify mesh→analytic-CAD reconstructions in OCCTReconstruct (the standing
mandatory-analytic-verification policy:
correctness must be proven by an independent OCCT MCP measurement, never a render). Two gaps blocked the check.

Gap 1 — no deviation / Hausdorff between a mesh and a shape

measure_distance returns minimum distance between two bodies. For a reconstruction overlapping its
source mesh that is ~0 and useless. What's needed to certify a reconstruction is the directed/symmetric
deviation field: max and RMS distance from mesh→solid (missing material / under-coverage) and solid→mesh
(over-extension), ideally with the worst-point location. This is the single most important verification
metric for measured reconstruction and there's no tool for it.

Proposal: measure_deviation(fromBodyId, toBodyId, samples?) → { meshToSolid: {rms,max,worstPoint}, solidToMesh: {rms,max,worstPoint} } (sample one body's surface, project to the other).

Gap 2 — import/measure tools hard-gate on validity, so an in-progress reconstruction can't be loaded

A reconstruction-in-progress is often a compound of loose analytic faces (not yet sewn into a valid
solid), and may contain a few invalid faces. Every normal entry point refuses it:

• import_file (STEP) → Error: Failed to write BREP … Shape is invalid or empty
• import_file with healOnImport:true → same
• read_brep (after converting to BREP) → same

I had to fall back to execute_script and compute Shape.bounds / subShapes(ofType:.face) by hand to
get any measurement. The whole point of verification is to measure the imperfect output, so the
analysis path shouldn't require validity. Concrete asks:

• import_file / read_brep: a force/allowInvalid flag that loads the shape as-is (compound/shell)
  without the BREP-rewrite validity gate, so compute_metrics (bbox/principalAxes), validate_geometry,
  and the proposed measure_deviation can run on it.
• compute_metrics should work on a loose-face compound (bbox/area at least).

Repro

• Source mesh: Fixtures/body_0_fixed.stl (valid)
• Reconstruction: 318-face loose-face STEP from reify … --nose-analytic whole trim (valid=false)
• import_file the STEP → "Shape is invalid or empty". The mesh imports fine; the reconstruction can't be loaded to be measured against it.

[gsdali]

Both gaps resolved in v1.9.0.

Gap 1 — mesh↔shape deviation. New measure_deviation(fromBodyId, toBodyId, deflection?, maxSamples?) → directed + symmetric surface Hausdorff: {max, rms, mean, worstPoint, samples} for fromToTo (over-extension) and toToFrom (under-coverage), plus symmetricHausdorff. Meshes both bodies, then exact point-to-triangle nearest-distance (KD-tree → incident triangles, Ericson) both directions. Available on both servers (PRs #46, #47).

Gap 2 — measure invalid / loose-face shapes. Root cause was the shape.isValid write-gate in OCCTSwift's Exporter (the analysis tools themselves load via the non-gating loadBREP and already work on invalid shapes). Fixed through the full chain:

• OCCTSwift v1.8.0 — Exporter.writeBREP(allowInvalid:) (feat(io): Exporter.writeBREP(allowInvalid:) to persist invalid shapes OCCTSwift#248)
• OCCTSwiftScripts v1.4.0 — load-brep / import --allow-invalid (feat(occtkit): --allow-invalid on load-brep / import (OCCTSwift 1.8.0) OCCTSwiftScripts#60)
• OCCTMCP v1.9.0 — allowInvalid on read_brep / import_file, both servers (feat(io): allowInvalid on read_brep / import_file (#41 Gap 2) #48)

With allowInvalid: true, an in-progress reconstruction (compound of loose analytic faces, possibly invalid) loads into the scene, and compute_metrics / measure_deviation / validate_geometry run on it. Verified with a deterministically-invalid bowtie face (rejected by default, loaded with the flag).

Closing as completed.