reflect: analysts see no trajectories for task ids that #309 maps, so LiveMathematicianBench training makes no edits #332

What happens. Since #309 (confine benchmark task output paths), rollouts write each task's outputs under predictions/<task_output_segment(id)>, so ids containing characters such as : are mapped to a safe directory name. skillopt/gradient/reflect.py still reads those outputs from the raw id, os.path.join(prediction_dir, tid, ...), in five places on main (84b36cd): conversation.json (line 141), target_system_prompt.txt (170), target_user_prompt.txt (182), codex_trace_summary.txt (195) and claude_trace_steps.txt (218).

LiveMathematicianBench ids have the form <month>:<number>, so every conversation.json lookup misses and fmt_minibatch_trajectories skips every item. Each analyst then returns no edits, and training runs all its epochs without changing the skill.

To reproduce. Train with configs/livemathematicianbench/default.yaml at 1a44050 (we used the claude_chat backend, Opus 4.6 as target and optimizer, on Linux with Python 3.12). Every step's analysts return 0 edits, and the skill stays at initial.md.

The fix we used. Read the outputs through the same mapping the rollouts use, falling back to the raw id:

from skillopt.envs.task_output import task_output_segment

def _task_dir(prediction_dir: str, tid: str) -> str:
    """Directory holding a task's outputs: the mapped segment rollouts write (#309), else the raw id."""
    mapped = os.path.join(prediction_dir, task_output_segment(tid, map_unsafe=True))
    if os.path.isdir(mapped):
        return mapped
    return os.path.join(prediction_dir, tid)

and use _task_dir(prediction_dir, tid) at each of the five reads. With it, the analysts get their trajectories again and training makes edits. In one LiveMathematicianBench run, test accuracy went from 46.4% for the starting skill to 74.4% for the best skill.

We're happy to open a PR with this change and a focused test, if that would help.