fix(navigation): project route flow with exact spherical matching

Flow samples were matched to the route through a longitude/latitude grid of
padded edge boxes, which claimed that any snap closer than the corridor would
be found. That is not true on a sphere: a great-circle edge bows away from the
box its endpoints describe. A route from [-1,80] to [1,80] rises about 165 m
poleward of latitude 80, so samples sitting essentially on the line — deviations
of a few nanometres — fell outside every candidate cell and the route reported
no flow at all. Widening the padding cannot fix this in general, because there
is no bound in degree space that holds near the poles, across the antimeridian
and for long edges at once.

Every sample now goes through the prepared spherical matcher, which prunes with
conservative great-circle bounds and evaluates survivors exactly, so the grid
and its run-local translation are gone entirely. Contiguous-run formation,
direction filtering, the backward-jump guard and pass selection are unchanged;
they simply sit on exact candidates now.

Overlap resolution was quadratic: it cut the route at every span boundary and
then scanned every span for every interval, about 50 million checks for five
thousand spans. It is now a boundary sweep with a max-heap ordered by severity
and then submission index, which is the rule the old scan expressed by taking
the first maximum it met. The ordering is explicit rather than inherited from
sort stability, so the tie behaviour is pinned rather than incidental.

One numeric consequence worth recording: along-route distance for a flow span
used to add a prefix computed with a 6371000 m earth radius to an offset
computed with Turf's 6371008.8 m, mixing two radii inside one value and leaving
flow distances on a different scale from navigation progress. It is now Turf
throughout, a shift of about 1.4 parts per million, and the two agree.

Span shape, ordering, direction, colours, thresholds, minimum span and merge
gap are unchanged. A seeded differential holds the new resolver against the old
one across 1440 span arrangements, and doubling five thousand spans to ten
thousand now roughly doubles the comparisons instead of quadrupling them.

Author: Medformatik