3-D cut-in-band: author the slit into the thin-volume assembly (the carve-based mid-plane cut cannot work)

[lmoresi]

The 2-D cut-in-band (ribbon + centre-line cut + free slit, #595) wants a 3-D counterpart: close-in fault resolution that is PLACED, not derived from a refinement cascade. Measured (~/+Simulations/place_route_health/ribbon_chain.py): the ribbon half already works — place_thin_volume embeds a fine layer-scale zone (3,388 cells at size 0.015) into an untouched coarse background (1,472 cells) in 3 seconds, the annular fill grading between the scales; no adaptation anywhere.

Cutting the fault along the ribbon's mid-plane AFTERWARDS fails structurally, not parametrically:

1. the mid-sheet's cavity shell needs cells held for the ribbon's skin (the sheet rim touches the SIDE skin; fixed by insetting the rim, which the split's tip rule requires anyway);
2. then the fill hits PLC segment/facet intersections wrapping the inset rim at matched fine scales — robust to clearance (0.8 and 1.0 tried) and width (2 and 4 layer-cells).

The 2-D solution never carved: #595 SNAPPED the cut onto the ribbon's existing rail vertices. The 3-D translation is to make the fault surface exist in the assembly BY CONSTRUCTION: mesh the thin volume as two half-slabs with assembly="fragment" (piece boundaries stay conforming), so the mid-plane is interior faces of the zone mesh; add a geometric labelling pass for those faces; split_along_label_3d opens them. No cavity, no fill, no PLC — the slit is authored, exactly as the 2-D spine was. Composition: coarse background + placed ribbon with built-in labelled mid-surface + split, ~16k cells for the benchmark fault against 49k-117k for the adapted alternatives.

Underworld development team with AI support from Claude Code

[lmoresi]

The authored slit WORKS end to end (place_route_health/ribbon_slit.py). Two half-slabs with assembly="fragment" keep the shared plane as conforming interior faces of the zone mesh; a geometric labelling pass selects them (inset from the ribbon rim per the tip rule) with one addition the first run taught us: ERODE all-rim faces from the selection until none remain — the split refuses a face with no interior vertex, and an arbitrary face-selection's corners can be all-rim (the placed-sheet path avoids this by construction). The erosion rounds the slit's corners slightly, which the tip rule is happy with.

Measured: ribbon placed in 4 s (6,758 zone cells at layer 0.015, coarse 1,472-cell background untouched); 710 slit faces labelled; split clean; frictionless contact solve 35 s, converged, machine-zero no-opening leak; peak slip 0.1509, consistent with the inset+eroded patch being ~0.34 against the 0.40 reference (crack slip scales with size: 0.85 x 0.178 = 0.151).

19,911 cells with fault resolution 0.015 — the same fault scale the wedge cascade needed 117k cells / 239 s for. The remaining quality tail (7 cells under q=0.05, min 0.0016) sits in the annular gap fill and the band transition, not on the slit.

Productionizing this = a mid_split=True mode on place_thin_volume (author the halves, label the interface, return the split-ready label) so users never hand-roll the labelling/erosion pass.

Underworld development team with AI support from Claude Code