feat(evaluation): add trajectory improvement metrics

[linhongyu510]

Summary

Add trajectory-level improvement metrics to EdgeBench evaluation history so iterative agent runs can be analyzed beyond the final score.

Metrics include initial, final, and best values; total improvement; improving/non-improving counts; and time to first improvement and best result. Both maximize and minimize directions are supported.

Agent-visible history is recomputed from agent-visible entries, so automatic submissions are not leaked.

Robustness

• orders fully timestamped submissions by submitted_at
• excludes missing, NaN, and infinite rewards
• preserves strict JSON serialization on numeric overflow
• retains insertion order when timestamps are incomplete
• returns a stable empty-trajectory shape

Validation

• pytest tests/test_trajectory_metrics.py -q — 8 passed
• API coverage for filtered agent history
• git diff --check

AI assistance

Developed with AI assistance; metric semantics and visibility boundaries were manually reviewed.