Archive QA round 4; document that rings merge disjoint outlines per section

SPEC.md

@@ -129,9 +129,10 @@ a rung doesn't reduce, so a stalled mesh never shows as several identical rungs.
 outline sampled at `samples` evenly spaced angles from the section's centre (the
 farthest crossing per ray), joined into bands with flat caps. Every band is one
 paper strip and every vertex is (ring, angle); the report lists each ring's height,
-centre and radii, so it can be checked by hand. One stack per object: branching or
-overhanging parts (a lamp's arm and shade) merge into one envelope, and heights with
-no section are skipped.
+centre and radii, so it can be checked by hand. One stack per object: each ray takes
+the farthest crossing, so disjoint outlines in one section merge (two legs become one
+band bridging the gap), branching or overhanging parts (a lamp's arm and shade) merge
+into one envelope, and heights with no section are skipped.
 
 ### Papercraft (rings)
 

feedback/vli-qa-feedback-2026-10-30.md

@@ -140,3 +140,21 @@ Seed-0 baseline reproduced bit-identically from the first session: uid `4459b90d
 **One new observation for round 4:** deep rungs can *spatially collapse*. The 384-face rung's `surface_area_m2` is 0.0017 vs 0.2233 on the 1,000 rung (~130× area loss) — the stall guard checks face count only, so a degenerate blob can pass as a legitimate rung, which misleads the papercraft/materialization use-case. Suggested guard addition: stop (or keep the previous rung) when a rung's area drops sharply, e.g. below ~50% of the previous rung's area. Page budgets look unaffected — the 400-face page cat keeps a normal ~0.3 m vertex spread — so this is a ladder-quality issue, not a page issue.
 
 Still open: `use` back-swap after a skip shifts indices; the UI visual click-test; P2-7 dictionary drift (now applies to the ForgeCAD SimReady source too).
+
+---
+
+## 7. Verification round 4 (2026-10-30, after the clustering/rings push) — all pass, one big addition
+
+| Item | Live result |
+|---|---|
+| Clustering fallback in `decimate` | ✓ `_clustered`/`_cluster_to` (grid-snap vertices + log-bisection on cell size) takes over when quadric stalls — the floor class of bugs is closed in principle, including SimReady CAD meshes (docstring cites a lamp stalling at ~21k faces on sliver triangles) |
+| Area-collapse guard | ✓ `_collapsed` implemented as suggested (extents + area vs previous rung; original excluded — welding double-sided faces legitimately halves area). The degenerate 384 rung is gone |
+| NEW `rings` rung | ✓ `rings(mesh, 8, 16)` — planar circumferences by ray casting; bands join into one paper strip each; profile attached to the rung metadata and returned in the report. On cat 138377: 256 faces, 130 verts, watertight, **area 0.2385 m² (sane)**, and the 8×16 profile reads as a sitting cat (base ring r≈0.098, leg gaps at y=0.045, tail kick: center shifts −0.108 at y=0.255, head top r≈0.014 at y=0.297). This is the SPEC's "materialization not built yet" gap under construction |
+| Reroll as queue rotation | ✓ `alternatives.append(chosen); chosen = alternatives.pop(0)` — seed-independent; matches the round-2 observation exactly (chosen `1f32bc` popped, skipped `by-nc` during settle, landed `138377`, old choice at the back) |
+| Clean generation reject | ✓ now raised in `grammar.next_tier` itself |
+
+Ladder shape note: for this cat the descent ends at 1,000 faces — the 250 request trips `_collapsed` (the clustered output lost >50% area) and the rings rung takes over as the low-poly representative. Healthy output, no degenerate rungs. Minor SPEC note: `rings` merges disjoint outlines per section (farthest-hit per ray), so a two-legs section becomes one band bridging the gap — fine for papercraft, worth documenting.
+
+**Still open (from the apple demo, unchanged by this push — confirmed by re-render):** grey SimReady objects (average the already-downloaded 256px thumbnail into `base_color`), dark no-sun scenes (`sun: null` → near-black render; suggest default daylight or a "what light?" question), cross-tier hint in questions (`q3 "which table?"` shows `alternatives: []` while 101 Objaverse tables sit one reject away), thumbnails as URLs so variants can be chosen by look (the 9 `food_produce_apple_*` variants are the real path to "a green apple"). Plus the minor `use`-back-swap check and the UI visual click-test.
+
+> Note (agent, 2026-10-30): the four apple-demo items above were addressed in 98ca16a, pushed after this round was written.