Workshop: two windows — Model Splitter and Model Fuser

[sswelm]

Summary

The Model Workshop is now two windows on one implementation: Model Splitter (Tools ▸ HAF ▸ Model Splitter) and Model Fuser (Tools ▸ HAF ▸ Model Fuser). ModelWorkshopWindow is abstract with one flag, Fusing; the two windows derive from it.

Shared by both: the file pickers, Probe, the filtered part list (verts, size, height, side, flat parts, Show only), the preview, "Find the mirror of …" and the ↑/↓ sweep.

	Model Splitter	Model Fuser
Row column	popup: – / Split / Delete	popup: – / ⊕A–Z / ✕ Delete
Keys	Insert = split, Delete = delete, Space = toggle split, – 0 Backspace = clear	A–Z = group, Delete = delete, – 0 Backspace = clear
List controls	Check all splittable, Uncheck all, Hide already-whole, Save/Load marks	Save groups, Load groups
Show only	checked, marked for deletion, islands, whole, skipped	in a group, not in a group, marked for deletion, islands, whole, skipped, every letter in use
Below the list	Plane cut, Split (and delete)	weld slider, Fuse (and delete)
Output suffix	_split / _cut	_fused

The one-step Generate (fuse AND split) went with the split. Chain the outputs instead: the ⊕ letters travel with a cut or split output (WorkshopRules.TransferLetters), and a fused output's sidecar names each shell under its group's letter (WorkshopRules.FusedOutputSidecarLines), so cut-then-fuse and fuse-then-cut both work. The Fuse button is twice as high and wraps, so nineteen groups' letters fit.

Marks persist — the Split checks and the deletion marks go to <source>.marks.txt beside the fuse sidecar (S = split, X = delete, the fuse sidecar's format and resolver), written by Save marks in the Splitter, by Save groups in the Fuser, and by every Split, Plane cut and Fuse; read at the first Probe of a file, and by Load marks / Load groups. Both windows share it, so a deletion marked in the Fuser reaches the Splitter. They are part of the pipeline, like the ⊕ letters.

Mark for deletion (both windows): a marked part loses its mesh in the output — its transform and children stay, the mesh data is left as an orphan the bake never reads — applied by Split, Plane cut and Fuse after their own work (GlbDisconnectedParts.RemoveMeshes, tested). A deleted part is neither split nor fused.

Also in this branch (Vehicle Lab, so you can test everything from one checkout): a Default reduce (%) dial — the catch-all tier that cuts every part still marked Default, exactly as Body parts are cut — and Preserve's dial moved below Detail's, the exception reading last. The rig script takes the new tier as a tagged argument (defaultreduce=@<file>|<pct>), so an older script and a newer Lab still run together, and a recipe that never touched the dial produces a byte-identical command line. Say the word if you would rather have this as its own PR.

Test plan

• bash tools/check.sh passes (853 tests; the Roslyn editor-compile step covers the window).
• Vehicle Lab: mark nothing, set Default reduce to 60, Verify → the projection shows the Default tier cut; Generate → the log's DEFAULT reduction lines name the parts. At 0 the command line is unchanged.
• Unity: Tools ▸ HAF ▸ Model Splitter shows checkboxes, plane cut and Split only; Model Fuser shows ⊕ popups, weld, Save/Load and Fuse only. Probe the Romanic in the Fuser: the sidecar's letters load, Fuse writes ss_romanic_fused.glb. Open that in the Splitter: the letters are still on the rows (hidden), a split writes _split and its sidecar.

🤖 Generated with Claude Code

[sswelm]

Self-review of 48d323b, fixed in the commit above:

1. Dead code. Generate(bool alsoSplit) kept its split branch after the Generate button went; unreachable. FuseMarked is the whole method now.
2. The mirror finder spoke the wrong language in the Cutter. It said "press a letter to mark it"; there Space checks the row, and a single-island part cannot be checked at all. Mode-aware now.
3. Three more Split-isms in the Fuser: the namesake warning, the Output GLB tooltip and the no-preview note. Mode-aware.
4. The Fuser dimmed its rows as "already whole" — the Cutter's concern. The Fuser's rows read "(N tris, N islands)" in the normal style.
5. The Fuser's intro promised the ⊕ letters travel with a fused output. A fused output carries none; they travel with a cut or split output, the only direction that needs them. Reworded.
6. Stale letters across windows. Each window reads the sidecar at Probe, so groups changed in the Fuser reach an already-probed Cutter only after a re-Probe, or its cut output carries the old letters. Documented in Editor-Tools.
7. Saved layouts. A layout holding the old Model Workshop drops that window once, the class being abstract now. Noted in the changelog.

Not changed, considered: both windows open on the same source each run their own Blender probe to the same preview FBX path — sequential, and each keeps its own instantiated preview, so no conflict.

🤖 Generated with Claude Code

[sswelm]

Addressed in the commit above: Fuse now writes the OUTPUT's sidecar as well — one line per fused group under the shell's own node index and name (every fuse Result names its shell), and the unchanged lines of a group that produced nothing. A fused output opened in the Cutter knows its groups; a cut shell's pieces inherit the letter; the Fuser welds them back if asked. Rule: WorkshopRules.FusedOutputSidecarLines, tested written-then-read-back; the parallel-fuse test checks each result's shell against the output's node table.

🤖 Generated with Claude Code

[sswelm]

Addressed in the commit above: a re-Probe in the Splitter now re-reads the sidecar. The letters kept across a re-Probe are the user's edits, which only the Fuser makes; the Splitter never edits a letter, so its letters are the sidecar's and a re-Probe reloads them — groups saved in the Fuser meanwhile reach the next cut, as the docs say. The Fuser keeps its edits as before, with Load groups as its explicit way back. Docs and changelog now describe the two behaviours precisely.

🤖 Generated with Claude Code

[sswelm]

The Split checks and the deletion marks now persist together in <source>.marks.txt, beside the fuse sidecar and in its format (S = split, X = delete, read by the same resolver, so a re-export that moves a part is resolved by name and index exactly as the ⊕ letters are). Written by Save marks in the Splitter, by Save groups in the Fuser, and by every Split, Plane cut and Fuse; read at the first Probe of a file, and by Load marks / Load groups. Both windows share the file.

🤖 Generated with Claude Code

[sswelm]

Critical self-review of the branch, findings and fixes.

Real defects, fixed

1. A delete-only Fuse cleared the ⊕ groups file. The Fuse button now fires with deletion marks alone, and the write path still called WriteFuseSidecar unconditionally; an empty write CLEARS the file (Save groups' documented "clear"), so a delete-only run silently dropped the groups saved earlier. It is written only when a group actually fused.
2. The two windows vanished from the layout on restart (found by the user, not by me). Unity binds a docked window's saved layout entry to the MonoScript whose FILE NAME matches the class; both derived windows were nested in ModelWorkshopWindow.cs, so the binding failed. ModelSplitterWindow.cs and ModelFuserWindow.cs now hold one class each, with their .meta.
3. The Default names file was written on every Generate, thousands of lines for an undecided model, even with the dial at 0. Written only when the dial asks.

Noise, removed

4. A BackSlash const had crept into the Vehicle Lab as scaffolding from the patch script that added the Default tier, where every other line spells the separator directly.

Considered, deliberately left

• RemoveMeshes strips one node's mesh reference, so a mesh shared by other nodes stays on those nodes. That is the right reading of "delete this part", and it matches what the fuse does to its sources.
• Marking a part for deletion does not grey it out in the preview. "Show only: Marked for deletion" is the way to find them, and the row label says so.
• Delete used to be one of the Fuser's clear keys and now means "mark for deletion"; the clear keys are – / 0 / Backspace. That is the requested behaviour, and the keys line states it.
• A sidecar that marks the same node both grouped and deleted resolves to deleted, in every load path.
• The Default tier cuts each part on its own, so a model with thousands of undecided shards costs real Blender time. The tooltip now says so rather than the code guessing a limit.

🤖 Generated with Claude Code

[sswelm]

The bridge-deck hole (user report) — a fusion defect, fixed in 50a334f.

Found by ray-casting the fused GLB from above and tracing every hole cell back to the source part that authored it: group D had 230 hole cells, 211 of them roof faces of Object_6 / Object_1774 that face UP in the source. The sheet holding them read double skin 67% twinned (74 twin in front / 8 behind, at 0.72 of reach) and inside-out score +0.55 — the twin rule reversed a sheet whose own score said outward. The 74 "twins" were the deckhouse's undersides a metre and more above: neighbours, not a skin.

Measured across the ship, every genuine double skin has its twin within a third of the old 1 % reach (hull skins 0.03–0.34, deckhouse skins 0.05–0.13, the fixture 0.20). The reach is 0.5 % now: every skin kept, every neighbour lost. Changing the precedence (score before twins) was tried first and broke the double-skin fixture, where the twins are right and the score wrong with the same signature.

holes from above (0.4 m cells)	before	after
Romanic, all groups	339	126
Romanic, group D	230	17 (authored so in the source: 35 there)
Teutonic, Fused_A	70	49
Teutonic, Fused_O / Fused_E	21 / 9	21 / 9

Teutonic per-group numbers unchanged. Each sheet's report line now names its parts and how far and how straight its twins lie — the two numbers that found this. 854 tests (new: a roof with a downward ceiling 0.7 % above it keeps its winding).

🤖 Generated with Claude Code

[sswelm]

Critical self-review, second pass — the whole branch diff against master, read after the reach change.

Fixed in the commit above (no behaviour change except 4)

1. Two comments glued onto the wrong line by patch scripts that matched a line's prefix: the IslandLines comment on the FusedNodeIndex line, the FuseLabels history on the Splitter's labels. Back where they belong.
2. The twin distance and straightness were labelled DIAGNOSTIC; they are report content, with the readings a skin gives (≤ 0.35 of the reach, ~0.9 straight) against a neighbour (0.7+) now spelled out.
3. A comment named a variable that never existed (skinReach).
4. RemoveMeshes stripped a node's skin with its mesh but not its morph weights — equally invalid on a meshless node. Gone too.

Checked and sound

• The reach change halves the enclosure veto's box slack as well: the nested-cavity fixture (2 cm on 10 m = 0.4 of the new reach) still holds, and the Romanic's hull shell reads the same 81 hole cells before and after.
• Deletion, split and group marks are mutually exclusive on every path — key, popup, and both sidecars (an X line wins over a letter in every load order).
• A delete-only Fuse leaves the groups file alone (fixed last pass); the explicit Save groups with no letters still clears it, as documented, and its status line says so.
• Per-group timings after the reach change: the 73,000-face hull group plans in 4.8 s.

Known and left

• A saved S mark is dropped silently if the part reads as one island at a different merge distance than it was saved at — it cannot be split then, and the status does not say.
• Insert on a whole part does nothing and says nothing.

🤖 Generated with Claude Code