turns out old code also had the same bug (just coincidance it was noticed after my change)
now boundry verts are tagged so edges connected to them are not seen is missing a tagged, adjacent edge.
this fixes [#30471]
Now create the rotate edge in advance and copy its customdata (before joining the faces).
This commit also fixes an annoyance where tryangulating faces could create duplicate edges.
On a user level, edge rotate now works better with multiple edges selected, it wont make zero area faces or rotate edges into existing ones.
With a single edge selected - rotate is less strict and will allow ugly resulting faces but still checks on duplicate edges.
API:
* BM_edge_rotate now takes a flag, to optionally...
** check for existing edge
** splice edge (rotate and merge)
** check for degenerate resulting faces (overlapping geometry, zero area)
** beauty - only rotate to a better fit.
... this allows it to still be used as a low level API function since all checks can be skipped.
* BM_edge_rotate() now works a bit different, it find the new edge rotation before joining the faces - exposed by BM_edge_rotate_calc().
* Added api call bmesh_radial_faceloop_find_vert() - Radial Find a Vertex Loop in Face
* optimize BM_face_other_edge_loop to do about half as many iterations for quad heavy meshes, with ngons the gain is much more since searching around the entire ngon when the edge already stores its loop is silly.
... also nicer in cases where edge has no face users it avoids a loop on all face corners.
* improve check to see if edge rotate can be done,
was checking if both edges verts have an edge count of 2, which is really a meaningless test since the verts can have stray edges connected and the result wont work right.
instead check if the next verts in both faces share a vertex.
* add utility function BM_face_other_vert_loop() which gets the next loop in a face.
* add convenience function BM_edge_face_pair() which returns 2 faces for edges that have exactly 2 face users. (saves ugly e->l->radial_next ... in code) and is more readable.
if blender freed the BMesh before python was finished (on exit editmode for eg), python would attempt to access the bmesh to clear python pointers in it.
- spelling - turns out we had tessellation spelt wrong all over.
- use \directive for doxy (not @directive)
- remove BLI_sparsemap.h - was from bmesh merge IIRC but entire file commented and not used.
* double default edge allocation size (double the number of verts/faces).
* CustomData_bmesh_init_pool was using allocsize & chunksize as the same variable. Now use type specific chunk size.
* bmesh copy and editmode conversion now allocate the BMesh mempool size needed for the entire vert/edge/loop/face arrays since its known already.
* name bmesh operator func's BMO_slot_buffer_* rather then BMO_slot_* since it wasnt obvious some only dealt with buffer, some both.
* more typechecks and asserts of BMO_ functions (I lost some time calling a map with a buffer function that failed silently).
* small speedup for extrude check - test if the edge is wire _before_ doign a hash lookup.
* add extra argument to BMO_slot_map_to_flag() to filter by vert/edge/face
* made BMO_slot_map_* / BMO_slot_buffer_* functions stricter with type checking.
added some missing functions too - which are not used yep but should be there for api completeness.
* CDDM_set_mloop
* CDDM_set_mpoly
* BLI_mempool_count
when making a quad from 2 edges - it was comparing the edge lengths to avoid making a bowtie quad. but this doesnt work in all cases, now compare normals instead.
