disinguish is_truncated and log stop_conditions

[rasdani]

• breaks down is_truncated into overlong_prompt and generation_truncation where it was lumped together previously
• adds verifiers stop_conditions to wandb logging

---

Note

Low Risk
Logging-only changes that add new metrics/columns without altering rollout generation or training behavior.

Overview
Adds stop_condition to the per-rollout results_df and emits new W&B metrics that break down truncation into generation truncation vs prompt_too_long.

Also logs the normalized rate of each observed stop_condition value (e.g. max_turns, has_error) to make rollout termination causes visible in monitoring.

^{Written by Cursor Bugbot for commit 48ff293. This will update automatically on new commits. Configure here.}

[cursor]

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

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Fixed: Generation truncated metric may never exclude prompt_too_long
  • Removed the broken generation_truncated metric that filtered on non-existent 'prompt_too_long' stop_condition value; the PR still logs actual stop_condition values via the dict comprehension.

Or push these changes by commenting:

@cursor push e3393b3f68

Preview (e3393b3f68)

diff --git a/src/prime_rl/orchestrator/orchestrator.py b/src/prime_rl/orchestrator/orchestrator.py
--- a/src/prime_rl/orchestrator/orchestrator.py
+++ b/src/prime_rl/orchestrator/orchestrator.py
@@ -667,10 +667,7 @@
             "is_truncated/mean": results_df.groupby("example_id").is_truncated.mean().mean(),
             "is_truncated/max": results_df.groupby("example_id").is_truncated.mean().max(),
             "is_truncated/min": results_df.groupby("example_id").is_truncated.mean().min(),
-            "stop_condition/generation_truncated": (
-                results_df.is_truncated & (results_df.stop_condition != "prompt_too_long")
-            ).mean(),
-            # Log rate of each stop condition (e.g. max_turns, prompt_too_long, has_error)
+            # Log rate of each stop condition (e.g. gibberish, repetition)
             **{
                 f"stop_condition/{sc}": rate
                 for sc, rate in results_df.stop_condition.dropna().value_counts(normalize=True).items()

_{This Bugbot Autofix run was free. To enable autofix for future PRs, go to the Cursor dashboard.}