fix(examples): redesign lod-decimate-chain rocket and show LOD density

README.md

@@ -866,7 +866,7 @@ Y-up and the differ check exits 9. Neighbor of
 </tr>
 <tr>
 <td width="46%" valign="middle">
-<a href="examples/lod-decimate-chain/"><img src="examples/lod-decimate-chain/preview.webp" alt="LOD decimate chain: three retro rockets side by side on a dark studio floor - cream body, red nose and fins, teal porthole - showing smooth LOD0, halved LOD1, and visibly crunched LOD2 with a hexagonal porthole" /></a>
+<a href="examples/lod-decimate-chain/"><img src="examples/lod-decimate-chain/preview.webp" alt="LOD decimate chain: three identical cream-and-red retro rockets with brass trim and a porthole on a dark studio floor, each under a dark triangle wireframe that coarsens left to right, labelled 4784, 2392 and 860 tris" /></a>
 </td>
 <td valign="middle">
 
@@ -875,9 +875,10 @@ Y-up and the differ check exits 9. Neighbor of
 One recognizable asset at three LODs via `DECIMATE` modifiers evaluated through
 the depsgraph. Asserts the reduction is non-destructive (the original datablock
 keeps its closed-form counts), each LOD's evaluated triangle count lands within
-5% of `ratio x base` (measured 0.0–0.44%), and the silhouette-critical bbox
-survives within 1e-3 — with the aggressive-ratio nose-tip collapse documented
-as the caught failure mode.
+5% of `ratio x base` (measured 0.00–0.13%), and the silhouette-critical bbox
+survives within 1e-3 — with a stacked Decimate and an oblique fin's shaved
+plate corner as the caught failure modes. A wireframe of each evaluated mesh
+makes the density drop visible.
 
 </td>
 </tr>

ROADMAP.md

@@ -165,7 +165,7 @@ Not committed; target list for the next content version. (v0.3.0 shipped the smo
 - ~~Light-linking collection witness~~ **SHIPPED** as `examples/light-link-studio/` — the API is `obj.light_linking` on the light OBJECT (`ld.light_linking` is an AttributeError on both versions); linked ratio 4.0x at projected-center luminance samples, unlink raises the decoy 244–251% with 0.0% hero drift in the same two-render check; EEVEE Next honors linking too (5.5x on 4.5.11, 5.9x on 5.1.2), Cycles pinned for deterministic samples
 - ~~VSE sequences-to-strips witness~~ **SHIPPED** as `examples/vse-cut-list/` — `.sequences` removed on 5.x (4.5 bridges to `.strips`), `new_effect` end kwarg `frame_end=` (4.5) vs `length=` (5.x), `frame_final_*` deprecated in favor of `left_handle`/`right_handle`/`duration`, TRANSFORM effect type removed, GAMMA_CROSS clamps to the source overlap, effect strips consume inputs only when stacked above them, same-scene scene strips render transparent; save/reload round-trip + tiny-render pixel witness. **5.2 COLOR `width`/`height` bake** (not in the python_api notes) taught in `skills/vse-python/`; the example `--check-pixels` path is the witness (set scene resolution before `new_effect`)
 - ~~glTF export round-trip witness~~ **SHIPPED** as `examples/gltf-export-roundtrip/` — `export_yup` bakes `(x,y,z)→(x,z,−y)` into vertex data with no node rotation (probed identical on 4.5.11 and 5.1.2), `export_apply` ships the evaluated mesh (one disk vertex per evaluated loop), TEXCOORD_0 is V-flipped on disk, per-triangle material bindings survive; exporter/importer RNA signatures byte-identical between 4.5.11 and 5.1.2 (guarded against future renames); `Mesh.calc_normals()` removal on 5.x surfaced during authoring
-- ~~LOD decimate chain witness~~ **SHIPPED** as `examples/lod-decimate-chain/` — Decimate COLLAPSE evaluated through the depsgraph is non-destructive (obj.data keeps closed-form counts), evaluated tris hit `ratio × base` within 5% (measured 0.0–0.44%), silhouette bbox survives within 1e-3 (measured 7.7e-6); a stacked Decimate halves the effective ratio, an aggressive 0.02 ratio collapses the nose tip — both caught failure modes
+- ~~LOD decimate chain witness~~ **SHIPPED** as `examples/lod-decimate-chain/` — Decimate COLLAPSE evaluated through the depsgraph is non-destructive (obj.data keeps closed-form counts), evaluated tris hit `ratio × base` within 5% (measured 0.00–0.13%), silhouette bbox survives within 1e-3 (measured 2.6e-10); a stacked Decimate halves the effective ratio, an oblique fin's plate corner is shaved at ratio 0.18 — both caught failure modes
 - ~~Vertex weight limit witness~~ **SHIPPED** as `examples/vertex-weight-limit/` — the 4-influence engine cap enforced via the data API (`v.groups` + `VertexGroup.remove` + renormalize); unit sums (measured 3e-8), pose preserved (4.9e-3), exact LBS from the mesh's own deform layer (2.7e-7), Root mount pinned
 - ~~Triangulate + tangent-space witness~~ **SHIPPED** as `examples/triangulate-tangents/` — `calc_tangents` aborts on any ngon (back cap must be an explicit fan); mikktspace matches the edge/UV-delta formula within welding tolerance on smooth fields (2.3e-6 measured); planar UVs on a cylindrical wall collapse tangents onto normals (dot 0.998); `MeshUVLoopLayer` handle dangles across `calc_tangents()` on 4.5 (471 phantom flips, silent exit 0) while the mikktspace math is byte-identical on 4.5.11 and 5.1.2
 - UV-handle lifetime snippet: re-fetch attribute/UV layers by name after any CustomData-reallocating call (`calc_tangents`, `VertexGroup.add`, modifier edits) — held handles dangle silently on 4.5, survive by luck on 5.1 (found authoring `triangulate-tangents`; **confirmed harder authoring `lightmap-uv-channel`: iterating a held `MeshUVLoopLayer.data` after edit-mode UV ops segfaults 4.5.11 headless, EXCEPTION_ACCESS_VIOLATION 5/5, survives on 5.1.2**)

docs/gallery/index.html

@@ -629,12 +629,12 @@ <h2><a href="export-preset-axis/">export-preset-axis</a></h2>
       </article>
       <article class="card" data-tags="depsgraph mesh">
         <a class="card-media" href="lod-decimate-chain/" tabindex="-1" aria-hidden="true">
-          <img src="assets/lod-decimate-chain-hero.webp" alt="Three red and white rockets labelled LOD0, LOD1 and LOD2, each more coarsely faceted than the last." loading="lazy" decoding="async" />
+          <img src="assets/lod-decimate-chain-hero.webp" alt="Three identical cream-and-red retro rockets with brass trim and a porthole, each covered in a dark triangle wireframe that grows coarser left to right, labelled 4784, 2392 and 860 tris." loading="lazy" decoding="async" />
         </a>
         <div class="card-body">
           <h2><a href="lod-decimate-chain/">lod-decimate-chain</a></h2>
-          <p class="teaches">A retro rocket at LOD0/1/2 via the Decimate modifier evaluated through the depsgraph. The check proves the reduction is non-destructive, the triangle count hits ratio x base within bounds, and silhouette-critical dimensions survive.</p>
-          <p class="witnesses"><span class="tag">witnesses</span> Evaluated-vs-original: the depsgraph carries the reduction while obj.data keeps the closed-form counts. A stacked Decimate halves the effective ratio (caught); an aggressive ratio collapses the nose tip (caught).</p>
+          <p class="teaches">A retro rocket at LOD0/1/2 via the Decimate modifier evaluated through the depsgraph, with a wireframe of each evaluated mesh. The check proves the reduction is non-destructive, the triangle count hits ratio x base within bounds, and silhouette-critical dimensions survive.</p>
+          <p class="witnesses"><span class="tag">witnesses</span> Evaluated-vs-original: the depsgraph carries the reduction while obj.data keeps the closed-form counts. A stacked Decimate halves the effective ratio (caught); a fin extreme on an oblique plate corner is shaved at ratio 0.18 (caught).</p>
           <a class="card-link" href="lod-decimate-chain/">View example<span class="sr-only"> lod-decimate-chain</span> <span aria-hidden="true">&rarr;</span></a>
         </div>
       </article>