feat: SAR sensor fusion — unmatched_sar_detections_30d (#84)

[yohei1126]

Summary

• Adds a new sar_detections table to store Sentinel-1 radar vessel detections (lat/lon/time, no MMSI)
• Implements unmatched_sar_detections_30d feature: counts SAR detections per vessel that occurred during an AIS gap and were spatially close to the vessel's last known position — a direct dark-vessel signal
• Wires the feature into the full scoring pipeline (feature matrix → Isolation Forest anomaly scoring)

Algorithm:

1. Match each SAR detection to AIS broadcasts within 5 km / 60 min → labeled matched
2. Unmatched detections are checked against per-vessel AIS gaps (> 6 h) using haversine distance (50 km attribution radius) to attribute to the most likely vessel
3. Count attributed unmatched detections per MMSI → unmatched_sar_detections_30d

All spatial joins are done in DuckDB (in-memory) with a bounding-box pre-filter + haversine verification.

Files changed

File	Change
src/ingest/sar.py	New — CSV/record ingestion into sar_detections
src/ingest/schema.py	Add sar_detections table + vessel_features column
src/features/sar_detections.py	New — compute_unmatched_sar_detections()
src/features/build_matrix.py	Wire SAR feature into feature matrix pipeline
src/score/anomaly.py	Add feature to ANOMALY_FEATURE_COLUMNS
tests/test_sar_detections.py	New — 6 unit tests
tests/test_schema.py	Update expected table set
tests/test_enhancements_cap_vista.py	Sync local ANOMALY_FEATURE_COLUMNS

Test plan

• test_empty_sar_table_returns_empty — no SAR data → empty result
• test_matched_sar_detection_not_counted — AIS-matched detections excluded
• test_unmatched_sar_during_gap_counts — gap + nearby SAR → counted
• test_unmatched_sar_far_from_vessel_not_counted — 500 km away → not attributed
• test_multiple_unmatched_detections_counted — 3 detections → count=3
• test_sar_outside_window_not_counted — 45-day old detection → excluded
• All 241 existing tests still pass

Closes #84

🤖 Generated with Claude Code