docs: publish M5 Max speed and compression research

[PhilipJohnBasile]

Summary

• publishes source- and receipt-backed research on M5 Max speed, compression, exactness, and dispatch behavior
• adds CPU-verifiable tools for dispatch, provenance, compressed-artifact comparison, and exactness auditing
• records negative results and non-comparable evidence explicitly; runtime defaults are unchanged
• keeps the research bundle repository-only and excludes it from sdist/wheel payloads
• provides a review map: 10 benchmark sources, 7 focused tests, 3 documentation files, 2 provenance files, 84 JSON artifacts including the manifest, and the package-boundary change

Verification

• 63 focused research tests passed
• 261 required public-contract tests passed (test_no_mlx_imports.py, test_public_cli.py, test_runtime_kpis.py)
• Ruff, format checks, Python compilation checks, and git diff --check passed
• deterministic manifest validated 23 sources, 83 receipts, 10 content-addressed references, and exact coverage of all 107 staged paths
• exact-index clean sdist and wheel builds passed scripts/fresh_venv_smoke.sh
• zero research paths were present in the built sdist or wheel
• immutable historical timing producer retained SHA-256 1578f6f8cb860be2e6dcb27ea08f1253b069555f0e83da3c71a451a0bcf98d33
• an independent frozen-diff review found no actionable blocker before publication

Evidence limits

• hardware: Apple M5 Max
• research dates: 2026-08-10 through 2026-08-11
• this is a build-selection research bundle, not a runtime-default change, model release, or production speed claim
• Qwen3.6-35B-A3B compressed-artifact timings use fixed token IDs and forward counts; expert assignments were not observed and may differ
• historical HumanEval NLL receipts are preserved but marked NOT COMPARABLE and have no decision authority
• measured, verified, published, and hypothesis claims are labeled separately in the main report
• the repository research bundle is intentionally excluded from Python distributions

Review entry points

• docs/research/mlx-m5max-speed-without-quality-loss-20260810.md
• benchmarks/results/mlx-m5-research-20260810/summary.json
• benchmarks/results/mlx-m5-research-20260810/bundle_manifest.json
• docs/research/mlx-lm-1709-exactness-audit-20260810.md

[PhilipJohnBasile]

Public CI follow-up: the fork did not auto-create pull_request runs, so I manually dispatched the existing release workflow against the exact PR head a204c6427292bf8b282b9a342f296f88824cd39e with publish_to_pypi=false. The macOS build-artifacts job passed; the PyPI publication job was skipped as intended: https://github.com/PhilipJohnBasile/MTPLX/actions/runs/31489988994

[PhilipJohnBasile]

@youssofal — GitHub would not retain a formal reviewer request on this fork, so I am requesting review here. The branch is evidence-only and changes no runtime defaults. The fastest review path is the main research note plus the content-addressed bundle manifest; the public macOS build/smoke run is linked above.