feat(monitor): upload native v1 episodes to Platform

[mrmoxon]

Summary

Restores the focused native Episode upload change from #3217 after it was reverted from the episode-first-class stack. This PR contains the same implementation tree as the previously reviewed #3217.

Stacked on #3206, which changes Prime-RL orchestration and monitors to carry complete first-class Episodes. Once that foundation reaches main, this PR can follow as the small upload layer.

Upload contract

• call verifiers' public build_samples() helper so eval and training uploads share one projection contract
• emit one Platform row per normal Episode
• preserve the complete Episode in info.native_wrapper
• retain info.native_trace_index and flat summary fields for legacy Platform consumers
• preserve every trace, including fixed-agent and traceless/error cases
• keep Prime-RL training fields such as run, step, environment, problem id, and advantage
• retain the 25 MiB per-trace fallback for oversized Episodes
• keep serialization best-effort so telemetry failures warn instead of aborting training

Stack

• Base/foundation: feat(orchestrator)!: make the episode the unit the orchestrator passes around #3206
• Original reviewed implementation: feat(monitor): upload native v1 episodes to Platform #3217
• This PR restores only the feat(monitor): upload native v1 episodes to Platform #3217 layer; it does not include the feat(orchestrator)!: make the episode the unit the orchestrator passes around #3206 foundation in its diff.

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit b36afc2. Configure here.}

[cursor]

Off-policy cancel one step early

High Severity

on_version_pending treats lag as (policy.version + 1) - policy_version, but the watcher already advances policy.version to the new step before this hook runs. That overstates lag by one and cancels groups that are still within max_off_policy_steps.

 

^{Reviewed by Cursor Bugbot for commit b36afc2. Configure here.}

[cursor]

Cancel emits wrong episode type

High Severity

drop_group emits vf.WireEpisode cancel markers, while emit_failed_episodes correctly builds Episode. Downstream code relies on Episode APIs such as train_run, rollouts, and env_name, so off-policy cancels can fail before markers reach the sinks.

Additional Locations (1)

• src/prime_rl/orchestrator/dispatcher.py#L586-L593

 

^{Reviewed by Cursor Bugbot for commit b36afc2. Configure here.}