Reward trend is a fragile two-point endpoint diff

[nsursock]

Problem

reward_delta_pct is a two-point endpoint diff (training.py:169-172): percent change from first to last finite sample of rollout/ep_rew_mean. RL reward curves are noisy; a single outlier final rollout flips the trend (reward_trend = +1/-1/0) and, transitively, the reward_delta_pct plausibility check.

Suggested fix

• Report a robust slope (Theil-Sen or MA-anchored mean-of-first/last-N) alongside the endpoint value.
• Keep the endpoint value for API compatibility but do not let it alone drive the verdict; add a CUSUM or at least an MA-delim check.

Impact

Agent-health "improving/degrading" conclusion is high-variance and can flip on one rollout.

Ref: src/dirty_fin_reports/simple/training.py:168-186; report.py:117-123