feat: cross-run regression detection

[jpleva91]

Problem

When a task that previously solved starts failing, there's no automatic alert. Regressions hide in the noise.

Acceptance Criteria

• Compare current run results against bench_history.json
• Flag tasks that solved in previous run but failed in current run
• Add "Regressions" section to sentinel report with task names and what changed

Hints

• evolve/history.py:detect_trends() already has improving/regressing logic
• Wire into sentinel_analyze.py after the per-task breakdown

---

Generated by /forge cascade