metrics(async): report selected policy provenance

[nanjiangwill]

Contract

Fully asynchronous training needs policy-provenance metrics for the data that actually reaches training. Missing version metadata must remain visible rather than becoming implicit zero-lag data.

Changes

• make avg_staleness and max_staleness cover only groups selected for training
• decompose selected-group staleness into post-generation staleness and the version span accumulated during generation
• report token-weighted staleness and selected-sample version coverage
• document that staleness is measured in published rollout weight versions, not necessarily optimizer steps

The staleness filter is unchanged: it still compares the current rollout weight version with the oldest numeric version in the group. stale_groups_filtered remains the rejected-group signal.

Cost

This reuses the WeightVersionSpan metadata already present on every sample. The existing group-version scan now also accumulates the newest version, token-weighted version, and sample coverage. It adds no RPC, payload field, per-token loop, CUDA synchronization, collective, or model forward.

Testing

53 passed in tests/fast/rollout/test_filters.py and tests/fast/rollout/test_fully_async_rollout.py
black: passed
isort: passed
ruff: passed

[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.

[Shi-Dong]

@claude review always

[claude]

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

[Shi-Dong]

Reviewed with local execution at c51b92c: 54/54 tests across both suites; the rescoped staleness metrics verified with an exact fail-before (base counts rejected groups, head does not); the generation-span/current_version=None finding was confirmed by probe on the previous head and the same probe now passes after the fix commit. Filter semantics are unchanged (delegation verified equal to the base implementation across nested/non-numeric/empty shapes).