Make the connected-net layout affordable on a dense patch
Laying a patch out as a connected net cost ~175 ms on a 42k-triangle patch,
against ~21 ms for the unwrap it works from, and the gizmo asks for it on every
preview, overlay and bake. Measured on a real project the grid behind it ran
~19 million triangle-pair tests per net, nearly all of them misses: a cell
holds every triangle whose box touches it, and a candidate really meets a
couple of them.
Keep a bounding box with each stored triangle and answer those misses with four
comparisons instead of a full intersection. The net drops to ~53 ms with
identical output - the seam metrics on the test project did not move by one.
Two further attempts were measured and dropped, and are recorded in the comment
so they are not tried again: a free-space pre-check per chart came out slower,
because a folded chart lands against the net by construction and the cells
under it are occupied anyway, and splitting the boxes into their own array for
locality lost more to growing two vectors per bucket than it gained.
Also pick a pair's fold line from the longest boundary they share rather than
whichever edge came first, and grow the net strongest-adjacency-first rather
than breadth-first by area. Only the fold a chart is reached by comes out
matching, so a chart claimed across a short boundary leaves the long one it
shared with its true neighbour torn.
texture_unwrap_dump reports an unwrap from a saved project - charts, their
topology, folded triangles, where the texture is discontinuous and how long
those seams are. All of the above was found with it, and it is what keeps a
claim about this code honest; reading the 3D view and guessing had produced
three wrong diagnoses in a row.
