Calibrated against a full-day legacy report for vehicle KDE 638J:
without this rule we collapsed 15 dispatcher-visible trips into 3
(the algorithm waited for explicit ACC_OFF + stationary fixes that never
came in the polled data — the device just stops reporting between trips).
New rule: if mid-trip and the next fix arrives >=5 min after the previous
one (but <=30 min, which is still long_gap), close the trip at the prior
fix with end_reason='nofix_stop'.
Validation:
638J full day: legacy 15 trips, mine 15 trips, perfect alignment
728K half day: legacy 33 noisy trips, mine 9 clean trips
(stop-and-go traffic still consolidates because fixes
keep coming, just slow/stationary — not silence)
Also commits scripts/simulate_trips_from_legacy.py which runs the same
state machine in Python against REPORTS/*.json dumps for offline tuning.
The CSV-based roster import (mig 15+16 and scripts/import_csv_roster.py)
merged vehicle rows that differed only by _Track / _CAM suffix, dropping
the active fleet count from 144 to 124. Reverting the whole thing.
Mig 17 in one transaction:
- Re-splits devices by parsed plate from device_name (same regex as
mig 14, preserving _Track as separate vehicle)
- Restores serve.fn_live_view to its v3 body (no d.driver_name/phone
refs that would break once the columns are gone)
- Drops the six CSV-only columns from domain.devices
- Deletes schema_migrations rows for the deleted 15/16
- Logs final counts via RAISE NOTICE
Apply on VPS: psql -f db/migrations/20260601000017_rollback_csv_import.sql
