The flip was my mistake. Measured against the service instead of guessed:
fetching a box and its northern half shows the half matching the top of the
whole, so the images arrive with their first row at the north edge and the
default texture coordinates were right all along. Both the base map and the
orthophoto go back to them.
Imagery of a given piece of ground does not change, so a fetched tile is now
kept under editor/cache/orto and the next session starts from it. Until now
the cache lived only in memory, which meant going back to a slow and flaky
service for ground that had already been fetched once.
Gaps can be fitted with a basket curve as well: the library's compound mode,
with clothoids at both ends and a chain of arcs between them. Arcs can be
added and removed; the closing one takes whatever deflection the earlier
ones leave, so its length is not offered for editing.
The location picker now shows the real topographic map instead of a
hand-drawn outline, fetched as one WMS image for whatever box is on screen;
the schematic border is gone and only the city list remains, to jump the
view somewhere recognisable. Under the plan itself a georeferenced map gets
the orthophoto, tiled through the layout library's own EPSG:2180 grid and
cache.
What a native build was missing is here: the download (WinHTTP, on worker
threads), the decode and the texture. Finished transfers are handed back on
the thread that drains them, so both the library's cache and the graphics
API stay on one thread. The imagery service turns down browser user agents
and fails the odd request even for a plain one, so requests go out as curl
and are retried a few times.
Starting a new map now drops the loaded scenery and comes back on empty
ground, instead of leaving the previous scenery's contents under a plan that
has nothing to do with them. The reload is deferred to the top of the frame:
taking the editor off the mode stack while it is drawing its own panels
would pull the ground out from under them.
