bench: adopt log_test_trajectory for sensitivity runs (drop Curve-B re-eval)

[davorrunje]

Context

benchmarks/_common/search.py:search has an optional log_test_trajectory: bool flag (added in PR #117) that records each trial's held-out test metric as the Optuna trial user-attr test_metric during the search — a pure diagnostic, never fed into selection. When present, benchmarks/_common/sensitivity_report.py:incumbent_test_curve reads that attr and needs no re-eval for Curve B. The committed studies predate the flag, so Curve B currently falls back to the (slow) incumbent re-eval.

Goal

Run dedicated sensitivity studies with log_test_trajectory=True so Curve B comes for free from storage, and retire the incumbent re-eval path as the default.

Where

• benchmarks/_common/search.py:search (flag already implemented).
• benchmarks/sensitivity.py (would consume stored attrs; no re-eval).
• benchmarks/_common/sensitivity_report.py:incumbent_test_curve (already prefers stored test_metric, falls back to re-eval).

Acceptance criteria

• A sensitivity study set is run with log_test_trajectory=True.
• Curve B for those studies reconstructs with # re-eval = 0.
• Docs note Curve B is now storage-only for runs using the flag.

References

• PR feat(bench): HP-search sensitivity curves (spec + plan + implementation) #117; spec docs/superpowers/specs/2026-07-15-hp-search-sensitivity-curves-design.md (follow-up 3).