fix(showcase): rebuild cart and stone-well joints, with eight new hygiene budgets

[TMHSDigital]

Quality pass over three items: two showcase pieces on Track A and one
check-only example on Track C. Both showcase pieces were the weakest in
the tree by budget coverage — 18 exit codes and two falsifiers each,
last touched in #171, before the joint-fit, named-support,
seat-conformance and mirror-symmetry convention classes existed.

Every defect below was found by rendering an inspection sheet to
.scratch/ (six orthographic views, two three-quarters, a clay pass, a
wireframe overlay, joint close-ups, a ground close-up) and looking at
the images. None of them is visible in the script.

---

1. showcase/cart — Track A

Defects found

#	Defect	View that revealed it
1	The iron tyre's inner cylinder and the felloe's tread are the same surface — 32 coplanar cross-shell face pairs, 16 per wheel, every one at 0.0000 mm separation. Rendered as grey speckle on the rim.	Clay three-quarter (speckle at 10 o'clock on the near wheel, in the same place in the clay, wireframe and material passes, so not shadow noise); confirmed by measurement
2	The felloe was a 16-gon. The chords read as a polygon at hero size, and the facet facing the key light rendered as a hard white plate.	Committed hero; top orthographic (white rectangles on both tyres); clay three-quarter
3	The iron read as black plastic, not metal — base colour 0.13 at roughness 0.32, darker and glossier than every sibling piece (hitching-post 0.16/0.48, wooden-ladder 0.18/0.48).	Committed hero; material three-quarter
4	Every run left a bdt_cart_*.glb in the repo root.	git status after a run

Defect 1 is the one that matters: it is a construction bug
(r_mid = RIM_OUTER + TYRE_T/2, radial_t = TYRE_T/2 makes
r_in == RIM_OUTER exactly), it shipped, and no budget could see it.

Fixes, and why they are structural

• Tyre: inner radius is now derived from the felloe's outer radius
  minus a named interference (TYRE_SEAT), and the tread from the
  felloe plus its own thickness. The band is hooped onto the rim and
  still owns the ground contact. A nudge to one radius would have
  reopened the moment RIM_RADIAL or TYRE_T changed; deriving both
  ends from the host surface survives that.
• Silhouette: WHEEL_SEGMENTS = 24, named, used by both rings and
  by the seat-depth binning, so the budget and the geometry cannot
  drift apart.
• Material: moved inside the range its sibling pieces already use,
  and the wear colour lifted off near-black so the noise varies the
  surface at all.

AABB is unchanged at (1.5393, 0.7486, 0.6400) — the tread radius
did not move.

New hygiene budgets

All recomputed from the generated mesh. Measured identically on
4.5.11 LTS / 5.1.2 / 5.2.1 LTS:

Budget	Measured
Coplanar cross-shell face pairs (hay-bale's combinatorics and constants, copied not imported)	0 (was 32)
Named supports — each tyre's own zmin, not just the AABB's	+Y 0.000000, −Y 0.000000
Tyre seat depth, banded, per angular station	[0.00400, 0.00400] over 48 stations
Wheel mirror in Y — position and extent	0.00000 mm

Falsifiers

Proven on all three binaries. Magnitudes sized to stay inside
BBOX_TOL so the AABB gate (8) cannot steal the failure.

Flag	Target budget	4.5.11	5.1.2	5.2.1
default	—	0	0	0
--skip-decimate	LOD1 ratio	9	9	9
--lift-z	AABB grounded	16	16	16
--flush-tyre	coplanar pairs (0 → 48)	15	15	15
--sink-tyre	tyre seat band (0.0090)	18	18	18
--float-wheel	named supports (+Y 0.003, AABB still grounded)	16	16	16
--skew-wheel	wheel mirror (6.000 mm)	19	19	19

--flush-tyre restores exactly the bug this piece was rebuilt to
remove, so the budget is proven against the real defect and not a
synthetic one.

Triangles: 2600 → 2856. Band re-fitted to 2760–2960, which is
narrower than the 2470–2900 it replaces (200 wide vs 430) and
centred on the measurement — a tightening, not a widening.

Inspection rounds: 2.

Remains (not fixed, reported rather than hidden)

• The far wheel is half-occluded by the bed and reads as floating in
  the hero.
• The shafts are constant-section sticks with no taper and no crossbar.
• The bed slats barely read; their ends show as small tabs under the
  bed frame.
• The wood is one flat value across the whole prop.

None of these is measurable without inventing a budget that would fail
on a correct model, so none became an assertion.

---

2. showcase/stone-well — Track A

Defects found

#	Defect	View that revealed it
1	The four roof posts stood at 0/90/180/270° while the roof's 4-gon cone is rotated π/4, putting its corners on the diagonals. The posts were under the midpoint of each eave: the roof's four corners cantilevered 0.80 m unsupported, and one post stood dead centre in the well mouth from every orthogonal direction.	Front orthographic showed three post columns where a four-post square shows two
2	Each post's bottom face landed exactly on the coping's top face — 4 coplanar cross-shell pairs, one per post.	Measurement; front orthographic (post feet sitting on a pale slab)
3	EAVE_HALF was derived from the post ring plus an overhang, giving a 1.64 m roof over a 1.08 m drum (1.52×). It read as an umbrella, or a bird feeder.	Clay three-quarter
4	The bucket hung at z=0.68 with only its rim level with the coping — invisible in the hero and in all six orthographic views. The piece's whole subject was down the shaft.	Clay three-quarter; front orthographic
5	Same .glb leak as the cart.	git status

Fixes, and why they are structural

• Posts move to the diagonals via a named POST_ANGLES, and the
  windlass, crank arm and grip turn with them, so the drum still spans
  an opposed pair of posts. Moving the roof instead would have been the
  nudge: the posts must be under the corners, not the corners moved to
  wherever the posts happen to be.
• Tenon: POST_BOTTOM is derived from the coping top minus
  POST_SEAT, and POST_TOP is held fixed, so nothing above the posts
  moves and the eave, peak and overall height are untouched.
• Roof: EAVE_HALF = R_OUTER + CURB_OUT + EAVE_CLEAR — derived from
  what the roof must cover, with the overhang falling out of it. AABB
  1.640 → 1.312 over a 1.08 m drum (1.21×).
• Bucket: BUCKET_Z derived from the coping top plus a named
  clearance, so it cannot drift back down the shaft when the masonry
  changes.

The tenon and bucket budgets measure against the coping top read off
the generated mesh (max z over the stone material), not against
CURB_TOP — an assertion that restates the builder's own constant
witnesses nothing.

New hygiene budgets

Budget	Measured (all three binaries)
Coplanar cross-shell face pairs	0 (was 4)
Named supports — all 12 bottom-course stones, not just the lowest	12/12, worst 0.000000
Post tenon depth, banded, per post	[0.01800, 0.01800] over 4 posts
Post-to-hip-corner plan offset, wrapped angular	0.00000 rad
Bucket clearance above the measured coping	0.10000 m

Falsifiers

Flag	Target budget	4.5.11	5.1.2	5.2.1
default	—	0	0	0
--skip-decimate	LOD1 ratio	9	9	9
--lift-z	AABB grounded	16	16	16
--stand-posts	coplanar pairs (0 → 4)	15	15	15
--float-stone	named supports	16	16	16
--shallow-tenon	post tenon band (0.0040)	18	18	18
--drop-bucket	bucket clearance (0.515)	18	18	18
--turn-posts	post placement (0.52360 rad)	19	19	19

Two falsifiers had to be re-aimed, which is the point of declaring
the target:

• --stand-posts seats the posts flush and returns them to the
  axis. Flush alone, at the corrected 45°, leaves no coping top face
  within COPLANAR_CENTRE_MAX of a foot, so it fell through to the
  tenon band (18) and witnessed nothing about z-fighting.
• --turn-posts rotates by exactly one curb facet (30°). At 0.12 rad
  the feet land on different coping geometry and the stone–wood gap
  gate steals it (observed exit 17, gap 0.01439 > 0.008); at a
  full 45° the crank grip swings past the eave and the AABB gate steals
  it (observed exit 8, bbox (1.3600, 1.3120, 1.7606)). One facet
  is the only rotation whose local seat geometry is identical by
  symmetry. Same trap stone-archway's --flat-arch fell into.

I also had to abandon a --twin-coping falsifier: duplicating a coping
block reuses its UVs, so the UV-overlap gate (7) fires before the
z-fight gate. Recorded here so the next run does not retry it.

Triangles: 9380, unchanged. Collider ceiling 320 → 380: the
convex hull of the resized roof measures 346. That is a budget moving
with a deliberate geometry change, documented at the constant.

Inspection rounds: 2.

Remains

• The top masonry course splays outward far enough that the well head
  reads as rubble rather than a dressed coping.
• The crank is a thin stub at hero size.

---

3. examples/ngon-triangulate — Track C

Contract frozen. No assertion, no exit code and no witnessed
behaviour changed.

A falsifier that falsified nothing

--skip-triangulate printed an error and returned 4 from its own
branch, before the repair and before the assertions it exists to
break. The flag announced a failure rather than causing one, and the
handling assertions never ran. It now skips bmesh.ops.triangulate and
lets those assertions catch the unrepaired mesh, which reports the real
measurement: triangulate handling failed ngons=1 tris=0 quads=4.

Exit codes, before and after

Identical on all three binaries:

Run	4.5.11	5.1.2	5.2.1
default	0	0	0
--no-dissolve	3	3	3
--skip-triangulate	4	4	4

Re-verification of the documented behaviour

Mesh.calc_tangents aborts on the constructed hexagon with a message
that is byte-identical on 4.5.11 LTS, 5.1.2 and 5.2.1 LTS:

Error: Tangent space can only be computed for tris/quads, aborting

TANGENT_ABORT = "tris/quads" is therefore correct and version-stable.
Default measurements also match on all three: ngons=1 loops=6 faces_before=5 tris=4 quads=4 loops_after=28 tangents=ok.

README

Adds the verified abort text, a falsifier table naming what each flag
breaks and the assertion it lands on, and 4.5/5.2 API links for
calc_tangents, bmesh.ops.dissolve_edges and bmesh.ops.triangulate
— it had no doc links at all.

---

4. showcase/README.md conventions

Seven conventions every future piece inherits, each from a defect above
that no existing budget could see:

1. A band is hooped onto its host, never set flush against it. The
   obvious spelling makes the band's inner cylinder and the host's
   tread the same surface.
2. A member is tenoned into its seat, never stood on it — and the
   seat is measured against the host surface read off the generated
   mesh, not the builder's constant.
3. Posts go under a roof's corners, not the middle of its eaves —
   with the corners taken from the pooled eave vertices, because
   pulling them from one shell picks a fascia board whose own extremes
   sit 90° off the corners it is nailed to (this cost me a false
   failure of 1.549 rad on a correct model).
4. A roof is sized from what it must cover, not from what carries
   it — and expect the collider budget to move with it.
5. The subject hangs where it can be seen, as a banded clearance.
6. Mirrored assemblies pair and compare, with the falsifier's
   displacement sized inside BBOX_TOL.
7. Segment counts are a silhouette budget, not a triangle budget —
   and re-fit the triangle band narrower, centred on the new count.

Plus this pass's falsifier flags registered in the shared list.

---

Gates

• Framing gate, 5.2.1, no deviation=: cart fill 0.806 (band
  0.70–0.90), margins 0.153/0.150/0.150/0.044, framing_ok; well fill
  0.861, margins 0.278/0.275/0.033/0.106, framing_ok.

• Contact sheets committed at
  docs/gallery/contact-sheets/cart-contact-sheet.webp and
  stone-well-contact-sheet.webp, each against the pinned set
  (armature-bend, damped-track-aim, bmesh-gear).

  Criterion	cart	stone-well	calibration range
  Mean luminance	46.50	49.03	43.89 – 77.66
  Saturation	0.1697	0.1369	0.1449 – 0.2386
  Stage darkness	pass	pass	—
  Wedge warmth	pass	pass	—
  Subject fill	pass	marginal — 0.861 in Y but 0.447 in X, the least-filling tile	—
  Thumbnail legibility	pass — reads as a two-wheel cart, rim now reads round	pass — reads as a wishing well, bucket visible	—

  stone-well's saturation is below the calibration floor (0.1369 vs
  0.1449): grey stone against one warm roof. Stated rather than
  massaged.

  The asset-sheet gate is scoped to asset-type examples; its
  pinned reference set is four examples, and every recent showcase pass
  (feat: add crate-stack and stone-archway showcase pieces #199, fix(showcase): hay-bale quality pass with five new hygiene budgets, and exit-pre-sidecar doc accuracy #205) shipped a contact sheet only. Followed that precedent.

• Determinism: both pieces run twice per binary, six runs each, all
  exit 0 with byte-identical measurements within each binary. Base
  triangles identical across all three series (2856 / 9380). The only
  cross-version difference is cart LOD2 (628 on 4.5/5.1, 616 on 5.2 —
  DECIMATE COLLAPSE), r2 0.2199 vs 0.2157, both inside the declared
  band.

• tests/check_falsifier_targets.py: 179 falsifiers across 28
  showcase pieces, passed.

• Generated gallery HTML re-read after build_gallery.py: card
  alt text and the witnesses callouts on both detail pages are
  complete and untruncated (the teaches.split(".")[0] class of bug
  from fix: gallery alts truncate on dotted API paths + version example prompt #68).

Proven by live run vs. inspection

Live run on all three binaries (.scratch/blender-4.5.11-windows-x64/blender.exe
reporting 4.5.11 LTS, blender-5.1.2-windows-x64 reporting 5.1.2,
blender-5.2.1-windows-x64 reporting 5.2.1 LTS): every exit code in
every table above, every measured budget figure, the two mis-aimed
falsifier exits (17 and 8) quoted with their messages, the
calc_tangents abort text, determinism, the framing gate, and the
temp-export cleanup.

Inspection only: the contact-sheet verdicts other than the two
measured numbers are my judgement of the committed composites, and the
"remains" lists are judgement calls, not measurements.

Inspection renders stayed in .scratch/, which is gitignored.

🤖 Generated with Claude Code