Make support-vertex renaming a fixed point and fix the #189 output regressions

[webdevred]

--transform was not a fixed point. updateSupportVertexName appended a side letter to a support name before assignNames ran, and assignNames was written to recognise its own output one character further along, so running the tool twice grew the name and nlsl1 became nlslsl. The pre-mutation is gone and assignNames now derives the whole name from the source name and the x coordinate, which makes a support prefix a fixed point of its own rule.

Two more things stopped it from settling once the names did. compareAV ordered support vertices by the name they currently have, which is the thing being replaced, so the first and second pass sorted on different keys. It now sorts on the prefix the vertex is about to be given. And breakVertices handed the comments and metadata in front of the next group back in reverse, because it reverses the finished tree but not the leftovers, which is why frame's // Middle side and // prefix group rl_m kept swapping places on every run.

Separately, #189 introduced three regressions in the example output that nobody caught because the fixtures were regenerated against the new behaviour. Support vertices reached the forest by two routes and keyed their entries differently, so they never collided and came out as two "Support nodes" sections with metadata stranded between them. Two OMap1 merges also dropped the incoming side's comments unconditionally, even when the existing side had none.

Frame's rear support node changes from rl_rsm to rl_sm. The side letter comes from the coordinate rather than from the name, and _r is a side letter as far as the transformation can tell. The ordinary nodes in that group already lost it, rl_r35 and rl_r38 have been coming out as rl_m0 and rl_m1 all along, so the two now agree rather than contradicting each other.

Outside the obvious scope: bfl_f and bfr_f in examples/jbeam/fender.jbeam are renamed back to bfl0 and bfr0. They were renamed to prove that letter-ending names survive a transform, but only the two node rows were changed, so every beam and triangle has been pointing at names that no longer existed. The property they were there to prove now has its own fixture under examples/regression_jbeam/.

Deliberately left alone: support classification still falls back to the node name, so a name ending in a letter becomes a support vertex regardless of how many beams reach it. That is wrong in principle, but it is not damage this branch caused, and it cannot be replaced until the connectivity threshold works. Measured across 3073 vehicle files, round(0.96 * n) exceeds the maximum possible degree n - 1 for every group under 25 nodes, which covers 97% of files, so dropping the name shortcut today would swap a wrong classification for almost no classification at all. Both halves have their own issue. #219 #220

Still draft, for two reasons. The example output wants a read from someone who knows jbeam, particularly the fender, which regains its side comments and its right-side group metadata. And suspension is not a fixed point yet, which is pending in the new spec: its three support nodes collapse to the same prefix, so the tie falls through to the metadata, and metadata is not carried across vertex tree boundaries. That is an older bug with its own issue and is not something this branch caused.