Two bugs left the terrain and models around the camera unbuilt on a big scenery:
- the first visual pass centred its section set on Global.pCamera.Pos, which the
driver hasn't positioned yet during load (reads 0,0,0) -> it built empty
sections at the origin and the spawn stayed bare. Centre on the sampled eye
(player / first vehicle) instead.
- the first (spawn) pass ran in the driver at a small per-frame budget on a live,
low-fps scene, so it crawled and never finished. Run it on the loading screen
(generous budget, nothing rendering), then hand to the driver once the spawn is
ready; the driver streams the rest as the camera moves.
Also drops the per-node capture index: capturing params/paths for a million nodes
cost more than the rescans it saved. Every cycle is now a plain scan-and-build of
the wanted sections (it finishes; a bake-time section index is the real cure for
the remaining first-pass scan time).
Result on tomaszewo: spawn now builds + renders (41k instances, ~hundreds of fps
once loaded). Remaining: load time is long (the 1M-node scan) -- next targets are
the getToken scan, the 380us/model build, and the 13s create_map_geometry finalize.
Adds load + frame profilers (behind WriteLog) used to find all of the above.
