feat(monitor): upload native v1 episodes without orchestration refactor

[mrmoxon]

Goal

Preserve complete native v1 Episodes in Prime-RL sample uploads without making Episode the orchestration unit.

What changes

• retain the original v1 Episode envelope as excluded, in-memory metadata when Env.run converts its traces to existing Rollout objects
• deduplicate that shared envelope in PrimeMonitor and serialize it through verifiers' existing build_samples() compatibility path
• preserve fixed-agent, failed, and other sibling traces in info.native_wrapper, while the existing trainable trace remains the flat summary for older Platform consumers
• keep legacy/group rollout paths working through a one-trace Episode fallback
• treat serialization failures as best-effort monitor failures rather than crashing training

What does not change

• Rollout remains the unit used by dispatch, scoring, training, metrics, and storage
• no Episode-first orchestration foundation is included
• no dependency on the earlier stacked Episode PRs
• the existing Platform upload endpoint and Parquet schema remain unchanged

Validation

• focused PrimeMonitor tests: 6 passed
• Ruff check and format check passed
• git diff --check passed

---

Note

Medium Risk
Changes telemetry serialization and Platform payload shape (info, episode grouping) but keeps training on rollouts; failures are best-effort and should not block training.

Overview
Prime Monitor now uploads native v1 Episodes to Platform instead of flattening each rollout with trace_to_sample, so multi-trace episodes (judge/solver/critic, failed siblings, etc.) stay intact on the wire while training still runs on rollouts.

Each v1 Env.run attaches an in-memory native_episode on every Rollout (full WireEpisode with orchestrator Rollout traces). Rollout gains an excluded native_episode field for monitors only.

log_samples rebuilds episodes via _rollouts_to_episodes (dedupe shared envelopes; legacy paths synthesize one-trace episodes), then build_samples projects rows and layers run/step/advantage/env_name as before. info comes from verifiers (e.g. native_wrapper, native_trace_index). Parquet build failures warn and skip upload instead of crashing training.

Unit tests cover multi-trace retention, legacy fallback, and serialization failure handling.

^{Reviewed by Cursor Bugbot for commit 208ba21. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

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

^{Reviewed by Cursor Bugbot for commit 8c4c303. Configure here.}

[mikasenghaas]

wait each rollout stores its back-stores its episode? this seems like a lot of redundancy and could lead to perf issues

imo this is a braoder smell of the episode-type not being fully native on prl orch yet. this is a known limitation @hallerite and me want to tackle this week. wondering if we should delay this pr until then

[mikasenghaas]

e..g do we even dedup the episodes if one episode makes multiple rollouts?

[mikasenghaas]

if we want to get in today, then we should try to pass a flat list of episodes directly to the prime monitor

[hallerite]

I think we should. the PR will be much cleaner after the refactor

[mrmoxon]

will leave it for now then