refactor: share trace_to_sample with verifiers in the prime monitor

[mikasenghaas]

Summary

• Replace the prime monitor's hand-rolled v1-trace → sample conversion in _rollouts_to_parquet_bytes with verifiers.v1.push.trace_to_sample — the single canonical "old format" converter, now shared with verifiers' eval uv run eval --push.
• Both producers (a training-run sample push here, and an eval --push / prime eval push in verifiers) now emit one identically-shaped sample record, so the platform renders them consistently and the conversion logic can't drift between the two repos.
• The RFT-only parquet columns (run_id/step/advantage/problem_id/env_name) are layered on here; the shared function owns the conversation/branch/trajectory extraction.
• Bumps the deps/verifiers submodule to the companion verifiers PR that adds trace_to_sample.

Behavior

Behavior-preserving: the _SAMPLE_SCHEMA and every column value are unchanged. The existing tests/unit/utils/test_prime_monitor.py asserts (problem_id/sample_id/completion/trajectory) still hold; I verified the refactored mapping produces identical num_input_tokens/num_output_tokens (still branches[-1]) and completion/trajectory content against those cases. The only nuance: messages now serialize with exclude_none=True (drops null fields like tool_calls: null) — smaller strings, same content.

Companion PR

Companion to the verifiers PR that introduced verifiers.v1.push.trace_to_sample (merged):
PrimeIntellect-ai/verifiers#1947

The deps/verifiers submodule is pinned to that PR's squash-merge commit on verifiers main (9c6b155e). Validated end-to-end with a 20-step reverse_text RL run on this branch (reward 0.78 at step 20, 0 errors).

---

Note

Low Risk
Refactor of monitor upload serialization with intended behavior parity; tests cover key columns; only minor JSON null-field omission differs.

Overview
Prime monitor sample uploads now build each parquet row from verifiers.v1.push.trace_to_sample instead of inline branch/message serialization, aligning training-run pushes with verifiers eval --push.

_rollouts_to_parquet_bytes still owns _SAMPLE_SCHEMA and RFT-only fields (run_id, step, advantage, problem_id, env_name, etc.). It maps example_id → problem_id, injects advantage on every trajectory branch, and reads token counts from the last branch. Empty trajectories are still skipped.

completion, task, reward, metrics, and timing come from the shared sample dict (JSON-encoded where before). Message JSON may omit null fields via exclude_none=True in verifiers—slightly smaller payloads, same semantics per PR notes.

^{Reviewed by Cursor Bugbot for commit 130bd1d. Bugbot is set up for automated code reviews on this repo. Configure here.}