docs: publish GLM-5.2 inference research closeout

[ZacharyZcR]

Summary

• publish the controlled inference-paper experiment matrix and failure ledger
• add reproducible placement, batching, prefix reuse, interference, disaggregation, and entropy tools
• retain fixed replay fixtures and the measured lossless split analysis

Key conclusion

For lossless single-stream GLM-5.2 decode on the tested 6x RTX 5090 / dual Xeon host, the remaining engine-only headroom is incremental. Future claims should reduce measured bytes per token or remove a measured whole-layer critical-path cost under held-out end-to-end evaluation.

Validation

• make -C c check: 292 Python tests passed, 18 skipped; native suite passed
• python3 -m unittest discover -s tests -p "test_placement_*.py": 8 passed
• CLI smoke checks for the included Python tools

Related: #537, #594, #452