- blend garbage-volume assertion checks the contract (kernel returned
more material than the 1000 input) rather than the macOS-specific
1036 value; Linux OCC reports ~1039 for the same r=6 fillet
- sca init shell-integration test pins SHELL=/bin/zsh; the .zshenv
target follows SHELL (or an existing .zshenv), and CI runs bash
P0 family 4 (sketch solve): Sketch.solve's two bare ValueErrors replaced
by structured diagnosis. The solver already names failing constraints;
they now surface as measurements (constraint/entity/DOF counts, failed
constraint values), inventory rows (failed / near miss = direct conflict
partners sharing target entities / ok), parameterized repair (conflict
pairs named; dimension-type constraints relaxed before geometric ones),
and a 2D matplotlib schematic following the diag_png pattern (implicated
entities orange, related purple, failing constraint ids badged at their
first target; extents include circle/arc/ellipse extremes; single view
- planes have no occlusion). New module sketch_diagnostics.py sits
beside sketch.py importing only errors (operators must never be imported
from the sketch layer). SimpleCADError subclasses ValueError so legacy
handlers keep working; raise_harness_error passes it through untouched,
so inspect/wire-promotion paths surface the payload intact.
Mechanism E (blend exit check): fillet/chamfer may only remove material
on convex edges - observed: all-edge r=6 on a 10-cube 'succeeds' with
volume 1036.35 > 1000. assert_blend_volume_monotonic names it at the
exit; BlendVolumeViolation leads what_happened (kernel refused vs kernel
lied) with probe diagnosis appended. Concave edges legitimately add
material (bolt underhead fill = quarter-circle deficit x circumference,
verified +1.28 mm^3) so the increase check requires a provably all-convex
selection via a new side-probe convexity test (n_a - n_b lands in
material iff concave; the bisector probe is 90/270-blind).
Tests: 8 sketch cases (text axiom, pixel-verified render, ValueError
compatibility, passthrough) + 3 mechanism-E cases (r=6 garbage caught,
legit fillet passes, concave-edge regression). Full suite 1848 green,
ty unchanged from baseline.
Three user-ruled corrections to the evidence renders:
- Highlighted edges are CRISP LINES, not marker cylinders: the render
core already had an unused highlighted_edge_groups line-actor channel
(4.5 px colored lines exactly on the edge); it is now exposed as the
public render_screenshot_rpath(highlight_edges=...) parameter and used
by every blend/QL diagnosis. No marker geometry covers the model.
- Evidence views are FOUR AXONOMETRIC positions (30 deg elevation,
azimuth 45/135/225/315), not one iso plus three orthographic panels
and not a single view: a highlighted feature hidden behind the model
in one iso is visible in another (the demo's back-most edge was fully
hidden in the previous single view).
- Edge ink is bolder in diagnostics (edge_width_scale 0.0035) and
highlight lines outweigh it (8 px). Boolean operand evidence drops
leader-line callouts: solid-tag anchors sit inside the solids and
occlusion re-anchoring can land on the face nearest the OTHER
operand, making the label point at the wrong body — legend colors and
the caption carry the mapping instead.
Diagnostic renders stay supersample=1 (fixed-width lines don't need
SSAA; keeps the 15 s watchdog comfortable). 1837 passed, 5 skipped.
The evidence renders were geometrically correct but visually useless:
marker tubes rendered in the same gray as the model. Evidence now drives
the render's highlight channel — tagged faces/solids get palette colors,
a legend, and callout labels:
- fillet/chamfer: the failing edge is a thick orange tagged tube
("failing edge" legend + callout). Two latent marker bugs fixed on the
way: a wrong relative-import depth and a missing TopoDS.Vertex_s cast
had silently killed every marker (and the edge_start/edge_end
measurements) behind the best-effort guard.
- boolean: the two operands render as tagged wrapper copies in
orange/purple — the user's own objects are never tagged (verified by
a pollution test).
- QL: matched faces highlight orange, up to three near-miss faces
purple, on a wrapper copy of the scope; matching is by kernel IsSame,
not Python shape ids (which differ across wrapper copies).
Highlight visibility is asserted at the pixel level (palette-color
sampled pixel counts), not by eyeballing. 1836 passed, 5 skipped.
Fillet/chamfer failures classify at raise time against the accumulated
degeneration catalog (repo history + OCC ecosystem): sliver edges in the
selection, tangent adjacent faces (dihedral probe via surface projection
— a G1 junction has no corner to blend for any size), blend size vs face
room (boundary-inclusive: r == room degenerates the strip, empirically
verified on a bore mouth where OCC accepts r=5 but fails at r=6 == the
6 mm annulus), partially selected smooth chains on periodic faces (the
boolean-seam split), crowded/mixed-convexity corner vertices, and the
2x-size tangency-critical interaction. Generic corner-blend conflict
goes last so specific explanations win; three same-convexity edges at a
box corner are routine and no longer shadow them.
QL cardinality violations now raise structured errors carrying the
selection-trust contract: an inventory (hit + near_miss with surface
parameters and axial spans, so split pieces of one surface stay
distinguishable in text), logical-surface grouping (faces on one
underlying quadric are one logical face the kernel split — pieces need
not be adjacent), parameterized repair naming the group, and one
rendered evidence view with boundary rods outlining the matched faces.
Symbol numbering follows the candidate enumeration so text and image
share one naming.
1835 passed, 5 skipped; ty clean on changed modules.