feat: witness n-gon, unapplied-scale, and coincident-vert pathologies

Gallery examples build clean geometry. These three synthesize the defects, pre-assert they exist, then check how hygiene/tangents, glTF export_apply, and export-crossing duplicates actually behave.

Signed-off-by: TMHSDigital <154358121+TMHSDigital@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

Author: TMHSDigital

.cursor-plugin/plugin.json

@@ -64,6 +64,7 @@
     "examples/attribute-domain-shear",
     "examples/bmesh-gear",
     "examples/car-mirror-symmetry",
+    "examples/coincident-vert-weld",
     "examples/collision-hull-proxy",
     "examples/color-attribute-wheel",
     "examples/compositor-glare",
@@ -90,6 +91,7 @@
     "examples/lod-decimate-chain",
     "examples/mesh-hygiene-audit",
     "examples/modular-kit-snap",
+    "examples/ngon-triangulate",
     "examples/parent-inverse-orrery",
     "examples/png-exr-alpha",
     "examples/prop-origin-transform",
@@ -103,6 +105,7 @@
     "examples/text-version-stamp",
     "examples/triangulate-tangents",
     "examples/turntable",
+    "examples/unapplied-scale-gltf",
     "examples/usd-export-evaluation-mode",
     "examples/uv-layer-grid",
     "examples/vertex-color-ao",

AGENTS.md

@@ -20,7 +20,7 @@ a `.cursor-plugin/plugin.json` manifest so the ecosystem drift checker
 classifies it as a `cursor-plugin`. This is content the AI loads when the user
 asks Blender questions or works on Blender add-ons in Cursor or Claude Code.
 
-The content base is 13 skills, 6 rules, 2 templates, 17 snippets, and 50
+The content base is 13 skills, 6 rules, 2 templates, 17 snippets, and 53
 examples (counts are CI-enforced against README.md and the manifest). The full
 inventory tables and per-item purposes live in `CLAUDE.md`. Example anatomy
 and authoring rules: copy `examples/bmesh-gear/`; the render look is specified
@@ -35,7 +35,7 @@ Blender-Developer-Tools/
   rules/<rule-name>.mdc          # 6 rule files
   templates/<template-name>/     # 2 starter templates
   snippets/<snippet-name>.py     # 17 standalone Python snippets
-  examples/<name>/               # 50 runnable smoke-gated examples (+ gallery.json)
+  examples/<name>/               # 53 runnable smoke-gated examples (+ gallery.json)
   examples/gallery_framing.py    # shared Layer 1 framing measurement (render path only)
   scripts/build_gallery.py       # generates docs/gallery/ (stdlib only)
   scripts/site/                  # vendored landing-page build (build_site.py + template)

CLAUDE.md

@@ -21,7 +21,7 @@ skills/<skill-name>/SKILL.md   - AI workflow definitions, 13 total
 rules/<rule-name>.mdc          - Anti-pattern rules, 6 total
 templates/<template-name>/     - Starter projects, 2 total
 snippets/<snippet-name>.py     - Standalone code patterns, 17 total
-examples/<name>/               - Runnable smoke-gated examples, 50 total (+ gallery.json)
+examples/<name>/               - Runnable smoke-gated examples, 53 total (+ gallery.json)
 scripts/build_gallery.py       - Regenerates docs/gallery/ from gallery.json (stdlib only)
 scripts/site/                  - Vendored landing-page build (Jinja2)
 docs/gallery/                  - Committed generated gallery pages + hero renders
@@ -83,7 +83,7 @@ v0.1.0: canonical object creation and deletion, depsgraph evaluated mesh, bmesh
 
 v0.2.0: Principled BSDF material, driver-with-custom-function via `driver_namespace`, application handler registration, shader node group with cross-version `interface` API, `foreach_get` bulk vertex read, version-branch skeleton, and USD export with `evaluation_mode='RENDER'`.
 
-## Examples (50)
+## Examples (53)
 
 Runnable scripts at `examples/<name>/`, each asserting a real API contract with
 deterministic checks (exit non-zero on failure) and optionally rendering a still via

README.md

@@ -18,7 +18,7 @@
 </p>
 
 <p align="center">
-  <strong>13 skills</strong> &nbsp;&bull;&nbsp; <strong>6 rules</strong> &nbsp;&bull;&nbsp; <strong>2 templates</strong> &nbsp;&bull;&nbsp; <strong>17 snippets</strong> &nbsp;&bull;&nbsp; <strong>50 examples</strong>
+  <strong>13 skills</strong> &nbsp;&bull;&nbsp; <strong>6 rules</strong> &nbsp;&bull;&nbsp; <strong>2 templates</strong> &nbsp;&bull;&nbsp; <strong>17 snippets</strong> &nbsp;&bull;&nbsp; <strong>53 examples</strong>
 </p>
 
 <p align="center">
@@ -627,7 +627,7 @@ portable path is `radius`.
 </details>
 
 <details>
-<summary><strong>Game asset pipeline</strong> — 9 examples</summary>
+<summary><strong>Game asset pipeline</strong> — 20 examples</summary>
 
 <table>
 <tr>
@@ -954,6 +954,51 @@ while **`BYTE_COLOR` is sRGB-encoded 8-bit** — 0.735 reads back
 Companion to [`color-attribute-wheel`](examples/color-attribute-wheel/) and
 [`attribute-domain-shear`](examples/attribute-domain-shear/).
 
+</td>
+</tr>
+<tr>
+<td width="46%" valign="middle">
+check-only, no gallery still — a hexagon on a cube is not thumbnail-legible
+</td>
+<td valign="middle">
+
+### [ngon-triangulate](examples/ngon-triangulate/)
+
+Synthesizes one 6-loop hexagon by dissolving a cube edge. Pre-asserts the
+n-gon exists (1 face / 6 loops / 5 faces), then `calc_tangents` aborts until
+`bmesh.ops.triangulate` yields 4 tris + 4 quads / 28 loops. glTF tri count
+is 12 either way. Inverse of [`mesh-hygiene-audit`](examples/mesh-hygiene-audit/).
+
+</td>
+</tr>
+<tr>
+<td width="46%" valign="middle">
+check-only, no gallery still — unapplied scale looks like modeled non-uniform size
+</td>
+<td valign="middle">
+
+### [unapplied-scale-gltf](examples/unapplied-scale-gltf/)
+
+Unit cube with object scale `(2, 1, 0.5)`. Pre-asserts unapplied non-uniform
+scale and local ±1 verts, then glTF `export_apply=True` still writes Y-up
+node.scale `(2, 0.5, 1)` with local POSITION. `export_apply` is modifiers
+only. Inverse of [`gltf-export-roundtrip`](examples/gltf-export-roundtrip/).
+
+</td>
+</tr>
+<tr>
+<td width="46%" valign="middle">
+check-only, no gallery still — coincident cubes look like one cube
+</td>
+<td valign="middle">
+
+### [coincident-vert-weld](examples/coincident-vert-weld/)
+
+Two cubes in one mesh: 16 verts / 8 unique / still manifold. Pre-asserts
+the duplicates, then glTF ships 24 tris / 48 positions / 8 unique.
+`remove_doubles` collapses to one cube. Inverse of
+[`degenerate-bevel-weld`](examples/degenerate-bevel-weld/).
+
 </td>
 </tr>
 </table>

ROADMAP.md

@@ -96,8 +96,11 @@ Not committed; target list for the next content version. (v0.3.0 shipped the smo
 - ~~Repeat Zone / For Each Element pairing witness~~ **SHIPPED** as `examples/gn-zone-iterate/` — `pair_with_output` is load-bearing (unpaired evaluates 0 verts); Repeat `8×(1+N)` with X-centers at `k×STEP`; For Each `8×P` with Z-centers at `i×STEP`; For Each main Geometry is a passthrough (6 verts); count-only is insufficient (`--no-offset` still 32 verts, one X-center)
 - ~~Geometry Nodes bundle round-trip~~ **SHIPPED** as `examples/gn-bundle-roundtrip/` — 5.x RNA `NodeCombineBundle` (`GeometryNodeCombineBundle` undefined on 5.2); closed form 8/6 + x `[0.5, 2.5]` + `bundle_mark=0.314159`; 4.5 SKIP via catalog `min_version` 5.0 (experimental flag off evaluates empty; `--force-run` hits 5.x RNA undefined)
 - Tighten inverted smoke canaries: assert the expected `[FAIL]` marker text (`skipped on ... should run` / `missing post-exit sidecar`) alongside wrapper exit 1, so a canary that dies for the wrong reason does not satisfy the gate
-- Falsification flags on shipped examples (Phases 6–8 `--skip-delete` / `--unpair-*` / `--bypass`, Phase 9 `--silent-handler` / `--wrong-text`) were proven local-Windows only — chat and `.scratch`, not CI history. Decide later whether a cron should exercise them or whether flags in the scripts are sufficient record.
 - ~~exit_pre sidecar witness~~ **SHIPPED** as `examples/exit-pre-sidecar/` — 5.1+ `handlers.exit_pre` writes `$BDT_SMOKE_SIDECAR`; harness `--expect-sidecar` + `sidecar_contains=exit_pre-ok` after process exit; 4.5 SKIP `min_version` 5.1; `--force-run` is AttributeError; `--silent-handler` / `--no-handler` missing sidecar; `--wrong-text` / `--write-in-main` / `--atexit-instead` wrong contents
+- ~~N-gon / unapplied-scale / coincident-vert pathology witnesses~~ **SHIPPED** as `examples/ngon-triangulate/`, `examples/unapplied-scale-gltf/`, `examples/coincident-vert-weld/` — two-stage pre-assert then handling; no gallery stills (defects not thumbnail-legible); `--no-dissolve` / `--identity` / `--no-duplicate` red on detection; `--skip-triangulate` / `--bake` / `--weld` red on handling
+- Overlapping / mirrored UV islands as a *shipped* pathology (not lightmap `--falsify`): SAT hits matching a constructed overlap, then glTF TEXCOORD survival. Deferred — `lightmap-uv-channel` already owns the zero-overlap gate.
+- Vendor CC0 asset fixtures (Kenney, Quaternius, ambientCG only — no aggregators) as `.glb` not `.blend`, with per-fixture provenance records mirroring the Free-Game-Dev-Assets frontmatter schema. Permitted only where the assertion is an invariant over the fixture rather than a measurement of it. Deferred pending a decision on repo weight and CI fetch policy.
+- Falsification flags on shipped examples (Phases 6–8 `--skip-delete` / `--unpair-*` / `--bypass`, Phase 9 `--silent-handler` / `--wrong-text`, Phase 10 `--no-dissolve` / `--identity` / `--no-duplicate`) were proven local-Windows only — chat and `.scratch`, not CI history. Decide later whether a cron should exercise them or whether flags in the scripts are sufficient record.
 - ~~Fleet Pages facelift + examples support~~ **RESOLVED differently (2026-07-03)**: the meta-repo migration was dropped — the fleet template only scaffolds new tools, and each tool's site evolves independently after that. This repo vendored the site build into `scripts/site/`, redesigned landing + gallery as the Blender-viewport system (see `docs/gallery/DESIGN_NOTES.md`), added the examples grid, nav link, and full hero stats locally. `scripts/build_gallery.py` and `examples/gallery.json` are now permanent, not a prototype awaiting lift-and-shift.
 - `modal-operators` skill -- `invoke` returning `RUNNING_MODAL`, the `modal()` event handler, modal cancellation patterns
 - `usd-pipelines` skill -- USD export options, `evaluation_mode`, instancing, the USD vs glTF tradeoffs

docs/technical-audit.md

@@ -72,6 +72,21 @@ on `bundle_items` are load-bearing. Vert count alone is not a witness.
 - Taught: `skills/geometry-nodes-python/SKILL.md`
 - Witness: `examples/gn-bundle-roundtrip/`
 
+### 2026-09-09 — glTF `export_apply` does not bake object scale
+
+RNA: "Apply modifiers (excluding Armatures) to mesh objects". Unapplied
+object scale is written to the glTF node, Y-up permuted
+`(sx, sy, sz) → (sx, sz, sy)`. POSITION stays the local mesh. Baking
+scale in Blender (`mesh.transform` + identity `obj.scale`) is what
+removes `node.scale` and puts world extents in POSITION. Same on
+4.5 LTS and 5.2 LTS; not a version split. `gltf-export-roundtrip`
+asserts no node scale because that crate is identity. Not listed in
+the python_api notes as a behavior change because the RNA name did
+not move.
+
+- Taught: `skills/depsgraph-and-evaluated-data/SKILL.md`
+- Witness: `examples/unapplied-scale-gltf/`
+
 ### Contrast — NodesModifier dict assignment (5.2, was in the notes)
 
 `mod[identifier] = value` raises `TypeError` on 5.2 (`id properties not

examples/coincident-vert-weld/README.md

@@ -0,0 +1,35 @@
+# Coincident vert weld
+
+Synthesizes two cubes occupying the same space as one mesh, then
+asserts the duplicates exist before asserting they cross glTF export.
+Inverse of [`degenerate-bevel-weld`](../degenerate-bevel-weld/)
+(coincidences from bevel pinch) and
+[`mesh-hygiene-audit`](../mesh-hygiene-audit/) (manifold on a clean
+solid). Neighbor of [`gltf-export-roundtrip`](../gltf-export-roundtrip/)
+(kit-bash face-plane welds on export).
+
+**Why this pathology:** coincident shells are still manifold (every
+edge borders 2 faces). Hygiene can pass. The engine trap is extra
+triangles on disk.
+
+**Pre-assertion (pathology exists):** **16** verts, **8** unique
+positions, **12** faces, **24** edges, valence all 2. `--no-duplicate`
+exits 3.
+
+**Handling (second axis):** glTF ships **48** loop-split positions,
+**24** tris, **8** unique. Vert count 16 is not enough; unique=8 is
+the construction axis; 24 tris vs 12 after `remove_doubles` is the
+export axis. `--weld` after the pre-assert exits 4.
+
+`remove_doubles` collapses to one cube (8/12/6), still manifold — not
+a 4-face-per-edge mesh. Same on 4.5 LTS and 5.2 LTS.
+
+No gallery still. Two coincident cubes look like one cube.
+
+## Run
+
+```bash
+blender --background --python coincident_vert_weld.py --
+blender --background --python coincident_vert_weld.py -- --no-duplicate
+blender --background --python coincident_vert_weld.py -- --weld
+```