fix(mobile): nav-start crash + stale-origin + missing recenter FAB

[cafca]

Summary

Three coupled bugs in the iOS navigation flow, all surfaced during a 20-minute real-device ride.

1. EXC_BREAKPOINT crash on nav start (GlitchTip BEEBEEBIKE-APP-4)

Int(p.durationRemaining * 1000) in Serialization.swift traps when ferrostar emits a non-finite duration on the first progress tick. GlitchTip captured the dict key duration_remaining_ms as the trap's value, pinpointing the line. Likely cause inside ferrostar: 0/0 from snapped speed=0 or a missing GraphHopper duration field. Root cause inside the Rust core not pursued — the FFI bridge shouldn't crash on degenerate input regardless.

Guarded:

• Int(durationRemaining * 1000) → finite + bounded check (NaN/Inf/overflow → 0)
• UInt16(courseDeg) → drop CLLocation's -1 sentinel and out-of-range values
• distance fields → fall back to 0 when non-finite

Defense-in-depth: maplibreToUserLocation in Dart also filters invalid bearing/speed/accuracy before they cross the boundary.

2. New routes started 20-30m back from current GPS

Geolocator.getLastKnownPosition() reads the system-wide cache, which lags MapLibre's own location stream by up to half a minute at cycling speed. Added a userLocationProvider (StateProvider<UserLocation?>) populated from MapLibre's onUserLocationUpdated — same source as the blue dot. Used as the preferred input when seeding nav, picking route origins (_handleMapTap, _navigateHome), and refreshing the preview after a reroute. Geolocator stays as a fallback for the cold-start case.

3. Recenter FAB and compass never appeared during nav

The awaitingFirstFix → following transition was wired to ferrostar's snapped_location stream, which only emits once a tick lands. On real devices that didn't happen for the entire 20-minute ride, so the camera stayed in awaitingFirstFix and the FAB never rendered. Now driven off the cached MapLibre fix at session start (synchronously) with onUserLocationUpdated as the fallback for the no-cache edge case. Renamed onFirstFix → onNavStart per design feedback ("first" was misleading mid-session).

iOS deployment target 13.0 → 16.0

Podfile already declared 16.0; pbxproj had been stale at 13.0. Pod-only sim builds let the mismatch slide, but SPM device-build resolution caught it (the ferrostar Swift Package requires .iOS(.v16)).

Test plan

• just test-mobile — 174 passing (3 new converter tests)
• just test-ferrostar-flutter-plugin — 21 passing
• flutter analyze clean
• flutter build ios --simulator --no-codesign clean
• Real-device ride: start nav, verify (1) no crash, (2) route originates from current GPS not 20-30m back, (3) recenter FAB + compass surface after panning the map mid-ride
• Confirm no new EXC_BREAKPOINT issues land in GlitchTip after a few sessions