Add self-contained TRELLIS.2 INT8 golden fixture

[benjipeng]

Summary

• add a self-contained miniature TRELLIS.2 checkpoint builder for regression tests
• run the production selective affine INT8 quantizer before exercising the real MLX inference pipeline
• compare reviewed tensor and GLB summaries against a committed golden manifest
• mark the end-to-end regression as heavy so default test runs remain fast
• document the explicit test command and manual rebaseline policy
• update the 8-bit model card with the verified 1024_cascade path

Why

The regression needs to detect quantization, checkpoint-loading, conditioning,
sampling, decoder, serialization, and GLB-export changes without retaining or
downloading the original TRELLIS.2 checkpoints. The generated source checkpoint
is used only as quantizer input; inference reads the temporary selective-INT8
bundle.

Impact

Default test runs continue to exclude the Metal-backed golden regression. Release
or inference-contract checks can opt in with -m heavy. Intentional output changes
require an explicit review and rebaseline of the committed manifest.

Validation

• .venv/bin/python -m pytest -m heavy tests/test_trellis2_golden_fixture.py -q
  • 1 passed in 64.58s after both full TRELLIS.2 weight bundles were removed
• .venv/bin/python -m pytest -q
  • 1175 passed, 10 skipped, 51 deselected, 3 xfailed in 27.22s