Add vLLM-Omni daily review knowledge

[hsliuustc0106]

Summary

• add a MiniMax H3 knowledge owner with task, shape, parallelism, checkpoint, audio fallback, and joint accuracy contracts
• add shared diffusion rules for FlashAttention capability selection and packed sequence boundaries
• add serving rules for bounded multipart artifact lifecycles and public CLI retirement
• add CI rules for test-selection closure and hardware-local oracle provenance

Why

A shadow daily review of the eight vLLM-Omni PRs merged on 2026-08-03 found durable contracts that were missing from InferMatrixCopilot knowledge. The review also showed that merged diffs alone are insufficient: critical conclusions were present only in review threads, and one merged PR targeted minicpm-challenge rather than main.

The batch keeps only executable owner-scoped rules. It does not persist raw PR evidence, incident pages, or one-off measurements.

Impact

PR review routing can now land MiniMax H3 changes in a dedicated model owner and apply focused rules for its task matrix, numerical shape, distributed topology, loading, audio, and accuracy behavior. Shared serving, diffusion, and CI reviews gain reusable gates discovered during the same intake.

No engine, adapter, playbook, or runtime behavior changes are included.

Validation

• python3 knowledge/tools/check_knowledge_tree.py
• python3 knowledge/tools/check_wiki_lint.py
• git diff --check
• PYTHONPATH=src python3 -m pytest -q test/test_knowledge_source.py test/test_adapters.py test/test_planner_playbooks.py — 23 passed
• full suite — 457 passed, 2 failures in unchanged code:
  • test_subprocess_tamper_defense: state was interrupted instead of failed
  • test_repo_neutral_core: config.py has three repository-specific literals against a ceiling of two

Follow-up

• run a separate source-sync update for configuration guides that still describe the retired public --stage-configs-path flag
• human-verify the upstream MiniMax H3 packed Ring, text-encoder subgroup, fused-loader, upload-limit, and joint audio-oracle findings

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.