[agrawalmilin9]

Summary

Implements GPS-based road learning using driver location data.

Closes #23

Changes

• Added GPS map matching and road-speed sample extraction.
• Added detection of followed and avoided road segments.
• Added validation to reduce false avoidance detection caused by GPS gaps.
• Added a background GPS learning worker.
• Added database models and Alembic migrations for GPS learning data.
• Saved planned routes for pickup and trip phases.
• Added GPS learning preview functionality and automated tests.

Testing

• Backend: 46 tests passed.
• Routing: 110 tests passed, 16 subtests passed.
• GPS detour integration test: passed using synthetic GPS coordinates without mocking map matching.
• Database migrations: verified at latest revision.
• Git diff checks: passed.

Notes

GPS learning is disabled by default and can be enabled using RUN_GPS_LEARNING_LOOP=true.

[aom12345]

This conflicts with main after #30. Please drop your own alembic.ini and migrations/ and add your migration under backend/migrations. It also doesn't yet deliver the issue: the learned data is never applied to the router (the test asserts refresh_telemetry isn't called), and the ROAD_LEARNING.md section is missing. Performance and correctness: nearest_segment scans every edge per ping on the event loop, so please add a spatial index and use asyncio.to_thread. Skipped or failed rides are never marked processed and starve newer ones. Speeds read up to 25% high because the full segment length is divided by a shorter time span. Planned-route computation runs even with the learning flag off.