Skill capability stays a hard filter, so a multi-skill tech (e.g. an
installer who also has the repair skill) can take those jobs. But the
specialist bias (skill-order rank) was too strong (6 virtual min/rank),
so a far repair-specialist beat a nearby installer-who-repairs. Lower
rank_weight to 2 → distance dominates; specialty only breaks near-ties.
Verified: repair job near a rank-3 multi-skill installer (4 min) vs a
far rank-0 specialist (14 min) — old weight picked the far specialist
(bug), new weight picks the nearby installer.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Real vehicle-routing optimization behind the Suggérer UI, reusing our
existing OR-Tools solver service (no new stack). Fixes the greedy's
structural limits (sector-splitting, no global optimization).
- roster-solver: new route_solver.py (OR-Tools Routing / VRPTW).
Minimizes real travel; skills = hard filter (VehicleVar ∈ allowed∪{-1};
SetAllowedVehiclesForIndex has a broken Span typemap in ortools 9.15);
on-site service time within each tech's shift window; optional per-job
time windows; unfittable jobs left unassigned (drop penalty) instead of
infeasible; specialist bias via skill order (per-vehicle arc cost).
New POST /route endpoint. Dockerfile now COPYs route_solver.py.
Unit-tested: skills respected, sectors consolidated, edge cases safe.
- hub: POST /roster/optimize-routes → proxies to solver /route.
- ops SPA: "⚡ Optimiser" strategy. Greedy buckets jobs into days +
placeholders, then the solver re-optimizes each day (assignment +
routes) among shifted techs; result maps into the same review dialog
(occupation bars, route map, swap/merge reused). Graceful fallback to
greedy if the solver is unreachable (never drops jobs). shiftWindowMin
derives each vehicle's shift from templates.
Phase 2 (later): OSRM/Mapbox road-time matrix for exact travel times.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
