fix: drain full trajectory batches after completion bursts

[original4422]

Summary

Keep completion bursts bounded while preserving rollout scheduling and failure checks between yielded batches.

refill rollout slots
check completed tasks (zero timeout when a full batch is buffered)
yield one full batch, then repeat
yield the final partial batch

Public signatures, skip behavior and exception handling are unchanged. Developed with AI assistance.

Evidence

• Before: eight immediately completed groups with batch_size=2 yielded [2, 6]; seven yielded [2, 5] and twelve [2, 2, 8]. Five burst regressions failed on main.
  After: batches stay bounded; all 12 public-API regressions and 3 existing localhost trajectory tests passed on macOS/Python 3.12. Event barriers also verify that completed pending failures propagate before another buffered batch and that new rollouts start before the backlog is drained.
• Complete uv run prek run --all-files passed with ty explicitly targeting Linux, matching the quality CI container, via an external config preserving all other repository settings. Default macOS ty has four pre-existing Linux-only attribute errors; ruff/format/lock passed. No hooks skipped. This does not claim Linux runtime, the full manual unit suite or GPU/backend validation.
• The separate exception-budget reset behavior addressed by Keep max_batch_exceptions for every batch in trajectory_group_batches #969 is unchanged.

Merge Danger

Door: two-way

Blast Radius: batching

Consumers receive bounded batches while rollout generation can overlap their processing.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T06:54:02.298349Z	a5fd7e2	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a5fd7e2524

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[original4422]

Hi @bradhilton, could you review this batching fix when you have time? The two automated timing concerns are addressed; the PR description includes the regression results and local validation limitations. Thanks!