[RL-Kernel] Add consistency audit harness

[inaniloquentee]

Summary

Implements #29 as the Phase 3 consistency audit harness on top of the Phase 1/2 stack.

• Adds a reusable vime.utils.consistency_audit path that composes Phase 1 consistency metadata validation with Phase 1 dlogp diagnostics instead of redefining either schema.
• Adds replay/debug manifests, batch-invariance replay case descriptors, per-rank manifest attachment in rollout DP splits, and compact result-cube records.
• Extends Megatron policy loss audit mode to use the shared harness and consume Phase 2 linear_logp runtime provenance when available.
• Keeps audit mode read-only: no resampling, regeneration, optimizer step, model mutation, kernel replacement, or distributed operator sweep is introduced here.

Stack and final merge path

This PR is intentionally stacked on:

• [RL-Kernel] Add linear_logp strict fallback checks #28 feature/issue21-linear-logp-strict-mode
• [RL-Kernel] Productionize linear_logp telemetry #27 feature/issue20-linear-logp-telemetry-phase2
• Phase 1 PRs [Core] Add RL-Kernel execution decision records #17, [Core] Add RL-Kernel consistency metadata fingerprints #18, Add audit-only dlogp diagnostics #19

The final target is still RL-Align/vime rlk-main. Keeping this PR based on #28 avoids duplicating the unmerged Phase 1/2 diffs during review. After the dependency stack lands in rlk-main, this PR should be retargeted or rebased to rlk-main and then merged there.

Validation

• py -3.13 -m compileall vime\utils\consistency_audit.py vime\backends\megatron_utils\loss.py vime\ray\rollout.py tests\utils\test_consistency_audit.py tests\utils\test_consistency_metadata.py tests\utils\test_consistency_metadata_rollout_integration.py tests\utils\test_dlogp_diagnostics.py tests\_unit_stubs.py
• py -3.13 -m pytest tests\utils\test_consistency_audit.py tests\utils\test_consistency_metadata.py tests\utils\test_consistency_metadata_rollout_integration.py tests\utils\test_dlogp_diagnostics.py tests\utils\test_rl_kernel_execution.py tests\test_rl_kernel_operator_adapter.py tests\test_rl_kernel_linear_logp_integration.py tests\test_megatron_argument_validation.py tests\test_cp_utils.py tests\test_loss_cp_invariance.py tests\test_dp_schedule.py tests\test_sample.py -q
• pre-commit run --all-files --show-diff-on-failure --color=always
• git diff --check

Fixes #29.

[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: 30b93cf2-dc1a-4795-b3e4-79203cf7a1ff

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/issue29-consistency-audit-harness

---

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.}