Strengthen regression test architecture

[benjipeng]

Summary

• separate routine, integration, real-asset, Metal, heavy, benchmark, and Torch-parity test roles
• keep one bounded routine integration guard for every supported model pipeline
• tighten TRELLIS.2 blockers, isolate generated outputs, and add a miniature MapAnything scene golden
• preserve strict cross-hardware contracts while using explicit invariants for ill-conditioned geometry and a narrow UV seam vertex tolerance

Why

The repository needs fast representative regression coverage as pipeline code grows, without requiring local production weights or allowing expensive checks to enter routine CI accidentally.

Validation

• local routine suite: 1181 passed, 61 deselected, 3 xfailed
• targeted cross-hardware golden checks: 3 passed, 1 deselected
• GitHub Actions Tests run 31844705145 passed on macos-15