feat(fuse): "Check mirrored parts" — an option for files that store mirrored parts already facing outward

[sswelm]

Summary

"There is an issue with Group A, it's partly transparent despite having all parts" (the Confederate frigate), then "make it an option".

glTF says a mirrored part (a negative-scale node, one side of a symmetric hull) renders with its winding reversed, and the fuse applies that to every mirrored part, which is exactly what the Teutonic needed. The frigate's file stores its mirrored hull half already facing outward, so the same reversal turned that whole side inward. Welded to the correct half in one island, the two cancelled every direction signal and the fuse left the island as it stood.

frigate group A, rays along the beam	toward −Y	toward +Y
fused, as today	94.8 % back-facing	0.2 %
the same parts unfused	0.1 %	0.1 %

The option, "Check mirrored parts", in the Model Fuser under the weld slider. Off by default. After welding, a plain part's winding is the reference, since no rule touched it. A mirrored part whose shared seams run the same way as its plain neighbours' is inside-out relative to them. The reversal is undone only on a clear verdict: at least three mirrored parts judged and nine in ten agreeing. Then every mirrored part follows, except one whose own seams disagree. Mixed evidence changes nothing. With the option off the fuse warns when a group would need it.

Measured with the real fuse on the real files

file, group	option off	option on
frigate A, hull	94.8 % back-facing	0.0 %
Romanic A, hull, from the beam	108 / 9 cells	8 / 5
Romanic D, deck, from above	161 cells	29
Romanic H (judged 1 pre-flipped against 2 standard)	—	byte-identical
Teutonic hull, eight other Romanic groups	—	byte-identical

The Romanic's 108 hull cells, taken this morning as authored in the source, turn out to be the same bug.

Two rules came out of the drills rather than the design:

• The file's convention. A per-part rule left the frigate's Object_961, which touches no plain part, on the reversal. It carried 2,443 of the 2,462 back-facing cells that remained. How mirrored parts are stored is a property of the exporter, so a clear verdict from the judged parts decides the rest.
• Clear verdicts only. Acting on a single part in the Romanic's group H, whose judged parts disagreed, made it worse, from 2,667 to 2,955 back-facing cells. Now mixed evidence changes nothing.

Developed and drilled in a separate worktree that the Unity project does not compile.

Test plan

• bash tools/check.sh (861 tests). New: a pre-flipped file fixed only with the option; a standard file byte-identical; mixed evidence unchanged; a lone mirrored part following the file's convention, its failure reproduced with the option off.
• With the option off, the frigate's output is byte-identical to master.
• Model Fuser on the frigate: fuse without the option and read the warning, then tick Check mirrored parts, fuse, and check the hull in the Vehicle Lab: both sides solid.
• The Romanic with the option on: the hull and deck as good or better, everything else unchanged.

🤖 Generated with Claude Code

[sswelm]

Confirmed and fixed in 516995e — thank you, it reproduced exactly as described.

A lap strip lies on its plate, stitched along one edge and facing the same way, so both faces walk the shared edge the same way on purpose. The consistency pass has handled that since 09-15 (the Teutonic's Object_8, 671 strips): same traversal with surfaces already facing the same way (normalised dot > 0.9) is consistent as authored. Step 3b counted it as a conflict. It now applies that same rule before counting — one lap rule in both places, not a second one — with a decided part about to be undone judged as it will be written.

Your case is the new regression test: three mirrored lap strips on plain plates — every face +Z with the option off, output byte-identical with it on, no verdict, no warning. It failed before the fix.

On the caveat that the real ships were not exercised: I re-drilled them, in a worktree the Unity project does not compile.

real file, option on	after the lap fix
frigate group A	byte-identical to the result that measured 0.0 % back-facing
Teutonic hull	identical to option off
all 11 Romanic groups	identical to the previous run — hull 108 -> 8 and deck 161 -> 29 gains stand

862 tests pass; gate passed on the push.

🤖 Generated with Claude Code