Make golden fixtures fast and self-contained - #8

Summary

• reduce the TRELLIS.2 selective-INT8 golden fixture from roughly 45 seconds to about 0.57 seconds by keeping the real SpatialKit path while using miniature export settings
• replace brittle floating-point byte digests with tolerant statistical summaries while retaining exact structural, integer, quantization, and topology contracts
• add a 65 KB real-weight-derived Pixal3D decoder patch with provenance, checksums, a reproducible fixture writer, and a Metal-backed native GLB replay
• document the distinction between synthetic full-pipeline coverage and real-weight-derived decoder/export coverage

Why

The previous TRELLIS.2 fixture spent almost all of its runtime in a production-resolution GLB export and its exact floating-point digests changed across otherwise consistent Metal runs. Pixal3D had fast synthetic full-pipeline coverage but no compact real-model boundary that remained usable after local checkpoints were removed.

Impact

Maintainers can detect major TRELLIS.2 and Pixal3D inference regressions without downloading or retaining the original model bundles. The two Metal-backed golden replays now complete in about one second combined, while ordinary CI retains fast provenance and contract checks.

Validation

• 1179 passed, 10 skipped, 52 deselected, 3 xfailed in the default suite
• TRELLIS.2 and Pixal3D heavy golden tests passed in three consecutive fresh processes
• measured heavy runtime: about 0.57 seconds for TRELLIS.2 and 0.33 seconds for Pixal3D
• git diff --check