[Sampling Replay 4/8] Capture native rollout sampling support

[nanjiangwill]

Stack position

Part 4/8 of the bounded sampling-support replay series. Merge after #3348. #3321 is merged in main.

Layer-only diff: nanjiangwill/miles@feat/sampling-replay-actor-scoring...feat/sampling-replay-rollout-capture

Review umbrella: #3354.

Summary

• Request SGLang v0.5.20 native realized sampling support for bounded training requests.
• Preserve SGLang support-normalized selected-token log probabilities.
• Validate complete per-token metadata and sampled-token membership.
• Cover standard, legacy, single-turn, and multi-turn rollout paths.
• Record environment and tool tokens with singleton support.
• Disable training-only capture during evaluation.

Correctness boundary

Miles does not reconstruct top-p or top-k support. SGLang is the source of truth for the realized support and owns physical support capacity and custom-logit-processor compatibility. Miles validates only transformations its trainer must reproduce, including request temperature and unsupported relative-logit changes.

This layer excludes session-server and agentic transport, which is handled by the next PR.

Validation

Focused tests cover request construction, native metadata parsing, abort behavior, ragged-support validation, forced tokens, multi-turn merging, and evaluation.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[nanjiangwill]

Closing this layer-specific review PR in favor of the single integrated sampling-replay PR #3354. The implementation and validation remain available there.