[Core] Add RL-Kernel consistency metadata fingerprints

[inaniloquentee]

Closes #10.

Summary

• Add �ime.utils.consistency_metadata for compact rollout consistency fingerprints, active-token accounting, requested-vs-actual provenance, audit warnings, and strict failures.
• Add Sample.consistency_metadata and carry metadata through Sample.to_dict/from_dict, default sample-to-train-data conversion, DP partition packaging, and batch-layout fingerprints.
• Annotate default vLLM/custom-generate samples when RL-Kernel consistency mode is enabled, and record dynamic-sampling keep/drop decisions without changing behavior when consistency is off.
• Add focused CPU tests for metadata schema, validation, provenance mismatch handling, zero-active-token reporting, Data Buffer preservation, rollout conversion integration, and Sample round-trip behavior.

Testing

• pre-commit run --all-files --show-diff-on-failure --color=always passed.
• python -m pytest tests/utils/test_consistency_metadata.py tests/utils/test_consistency_metadata_rollout_integration.py tests/test_sample.py -m unit passed: 31 tests.
• python -m pytest tests/plugin_contracts/test_plugin_generate_contracts.py tests/plugin_contracts/test_plugin_rollout_contracts.py tests/plugin_contracts/test_plugin_path_loading_contracts.py passed: 24 tests.
• python -m pytest tests/utils -m unit ran 135 selected tests: 133 passed, 2 failed in local vLLM argument tests because the installed/stubbed vLLM surface does not expose --vllm-tool-call-parser; this is unrelated to the consistency metadata changes.
• python -m pytest tests/plugin_contracts/test_plugin_generate_contracts.py tests/plugin_contracts/test_plugin_rollout_contracts.py tests/plugin_contracts/test_plugin_path_loading_contracts.py tests/test_vllm_rollout.py -m unit showed 30 passed and 3 existing ests/test_vllm_rollout.py local failures unrelated to this PR.

AI assistance

Implemented with OpenAI Codex assistance; commit includes Co-authored-by: OpenAI Codex codex@openai.com and DCO signoff.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 5b5a873c-71c8-49d9-9ba2-8d3ae13bebbe

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch feature/issue10-consistency-metadata

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}