[M5] Build AI 3D evaluation harness for regression samples

[weiuou]

Goal

Create an offline evaluation process for comparing Meshy, TripoSR, Hunyuan3D, TRELLIS, and procedural fallback against the existing regression sample manifest.

Scope

• Add a documented evaluation harness that reads docs/qa/regression-sample-suite.json positive samples.
• Record provider success, mesh parseability, vertex/face counts, paperability repair, decimation, unfolding, export success, page count, part count, and fallback status.
• Add a result template for manual ratings: subject similarity, cutability, and whether the sample should enter manual paper QA.
• Do not require actual sample binaries in the repo.

Acceptance Criteria

• Evaluation can be run in dry-run/mock mode in CI or local tests.
• Output format is stable enough to compare providers across runs.
• Documentation explains how to run Meshy API and local model candidates.