feat: structured failure classifier output format

[jpleva91]

Problem

sentinel_analyze.py produces markdown reports but no machine-readable output. The evolve pipeline needs structured JSON for automated decision-making.

Acceptance Criteria

• sentinel_analyze.py outputs a run-NNN-classified.json alongside the markdown report
• JSON includes per-task: task_name, reward, failure_type, tests_passed, tests_total, agent_duration
• Format matches bench_history.json schema for easy merging

Hints

• See bench/sentinel_analyze.py:analyze_job() — the task data is already structured internally
• See bench/evolve/classify.py — the classifier is already integrated
• Just need to add JSON serialization at the end of analyze_job()

---

Generated by /forge cascade

[jpleva91]

@claude Please implement this issue. Read the full codebase first to understand the architecture, then implement per the acceptance criteria. Run go test ./... and go build ./... to verify.