feat(swarm): GeoBridge Module #68 — Signal #21 server-side geo proximity (80 tests)

[0xultravioleta]

What

GeoBridge adds server-side geo proximity intelligence to the EM swarm coordinator. This is Module #68 and the server-side counterpart to AutoJob's GeoProximityEngine (Signal #21).

Why

Signals #1-20 optimized worker routing on behavioral/reputation/motivational dimensions but completely ignored geography. For physical tasks (photo, photo_geo, video, measurement), a worker 2km away is fundamentally better than a worker 7,000km away — regardless of reputation.

Signal #21: Geo Proximity

Four sub-signals (weights sum to 1.0):

• Haversine Proximity (50%): exp(-dist_km/10) decay from worker's last known location
• Territory Intelligence (30%): Historical completions in nearby 1km² grid cells
• Commute Willingness (10%): Worker avg travel distance vs 5km market median
• Temporal Clustering (10%): Active hours overlap with task's time window

Task type routing:

• Physical tasks (photo, photo_geo, video, measurement, etc.) → full geo signal
• Digital tasks (text_response, screenshot, etc.) → geo_bonus = 0.0

Geo bonus range: ±0.08 max contribution to match_score.

GPS Extraction

Handles all 4 EM evidence formats (gps nested, flat, metadata, direct columns).

Tests: 80 passing

Files: mcp_server/swarm/geo_bridge.py (633 LOC), mcp_server/tests/swarm/test_geo_bridge.py (80 tests)