fix: clean stale rollouts on fresh runs

[mikasenghaas]

Summary

• If a run is started from scratch in an output_dir that still contains stale rollouts or broadcasts from a previous run (but no checkpoint), the trainer would consume those stale rollouts and the orchestrator would see a negative async level because the trainer appears ahead of it.
• clean_future_steps previously only ran when resuming from a checkpoint. Now it also runs from step 0 (i.e. resume_step=-1) when training from scratch, wiping every existing rollout/broadcast step directory before training begins.
• Also updates clean_future_steps' docstring to document the new resume_step=-1 semantics.

Repro

uv run rl @ examples/reverse_text/rl.toml --max-steps 3 && rm -rf outputs/checkpoints && uv run rl @ examples/reverse_text/rl.toml --max-steps 3

---

Note

Medium Risk
Moderate risk because it changes startup behavior to delete existing rollout/broadcast step directories, which could remove data if an output_dir is reused unexpectedly. The logic is localized to pre-run cleanup and should reduce mis-synchronization issues when starting from a dirty directory.

Overview
Pre-run cleanup now always invokes clean_future_steps: when resuming it deletes steps after the resolved resume_step, and when training from scratch it calls clean_future_steps(..., -1) to wipe all existing rollout/broadcast step directories in the output_dir.

Updates clean_future_steps’s docstring to document the new resume_step=-1 semantics for fresh runs.

^{Reviewed by Cursor Bugbot for commit 8fb4011. Bugbot is set up for automated code reviews on this repo. Configure here.}