intersect_geom! in postrender_ops silently drops entities

[parrangoiz]

intersect_geom! in postrender_ops silently drops entities from a multi-entity 2D physical group when intersected with a 3D volume. This isn't a bug, as OCC intersection ops are meant to be run on fragmented/conformal models AFAIK, but it's certainly a footgun as any user who schedules a mixed-dimension intersection in postrender_ops is in general not going to get the right output. Here's a minimal reproduction with DL v1.13.0:

using DeviceLayout
using DeviceLayout.PreferredUnits
using DeviceLayout.SolidModels
using DeviceLayout.SolidModels: difference_geom!, intersect_geom!, extrude_z!

cs = CoordinateSystem("repro", nm)

# Layer 1 (→ PG :metal): ground plane (with hole) + island
ground = DeviceLayout.difference2d(centered(Rectangle(100μm, 100μm)),
                                   centered(Rectangle(40μm, 40μm)))
place!(cs, ground, :metal)
island = centered(Rectangle(20μm, 20μm))
place!(cs, island, :metal)

# Layer 2 (→ PG :substrate): chip outline (becomes substrate)
chip = centered(Rectangle(120μm, 120μm))
place!(cs, chip, :substrate)

# Layer 3 (→ PG :box): chip outline (becomes bounding volume)
place!(cs, chip, :box)

# --- Case A: intersect inside postrender_ops (BEFORE _fragment_and_map!) ---
sm_a = SolidModel("repro_A"; overwrite=true)
render!(sm_a, cs;
    postrender_ops=[
        (:substrate, extrude_z!, (:substrate, -50.0μm)),
        (:vacuum, extrude_z!, (:box, 50.0μm)),
        (:interface, intersect_geom!, (:metal, :vacuum, 2, 3)),
    ],
)
n_a = SolidModels.hasgroup(sm_a, :interface, 2) ?
    length(SolidModels.entitytags(sm_a[:interface, 2])) : 0

# --- Case B: intersect AFTER render! (AFTER _fragment_and_map!) ---
sm_b = SolidModel("repro_B"; overwrite=true)
render!(sm_b, cs;
    postrender_ops=[
        (:substrate, extrude_z!, (:substrate, -50.0μm)),
        (:vacuum, extrude_z!, (:box, 50.0μm)),
    ],
)
result_b = intersect_geom!(sm_b, :metal, :vacuum, 2, 3)
n_b = length(result_b)

println("Case A (intersect in postrender_ops): $n_a entities")
println("Case B (intersect after render!):     $n_b entities")
println("Expected: both should return 2 (ground plane + island)")

[gpeairs]

Example reproduces with just 2D entities:

using DeviceLayout
using DeviceLayout.PreferredUnits
using DeviceLayout.SolidModels
using DeviceLayout.SolidModels: difference_geom!, intersect_geom!, extrude_z!

cs = CoordinateSystem("repro", nm)

# Layer 1 (→ PG :metal): island
island = centered(Rectangle(20μm, 20μm))
place!(cs, island, :metal)

# Layer 2 (→ PG :substrate): chip outline (becomes substrate)
chip = centered(Rectangle(120μm, 120μm))
place!(cs, chip, :substrate)

# Layer 3 (→ PG :box): chip outline (becomes bounding volume)
place!(cs, chip, :box)

# --- Case A: intersect inside postrender_ops (BEFORE _fragment_and_map!) ---
sm_a = SolidModel("repro_A"; overwrite=true)
render!(sm_a, cs;
    postrender_ops=[
        ("interface", intersect_geom!, (:metal, :box, 2, 2)),
    ],
)
n_a = SolidModels.hasgroup(sm_a, :interface, 2) ?
    length(SolidModels.entitytags(sm_a[:interface, 2])) : 0

# --- Case B: intersect AFTER render! (AFTER _fragment_and_map!) ---
sm_b = SolidModel("repro_B"; overwrite=true)
render!(sm_b, cs;
    postrender_ops=[
    ],
)
result_b = intersect_geom!(sm_b, :metal, :box, 2, 2)
n_b = length(result_b)

println("Case A (intersect in postrender_ops): $n_a entities")
println("Case B (intersect after render!):     $n_b entities")
println("Expected: both should return 1 (island)")

Case A (intersect in postrender_ops): 0 entities
Case B (intersect after render!):     1 entities
Expected: both should return 1 (island)

If :metal is not fully contained in :box (e.g., the box rectangle is not centered), then we get the expected result in both case A and B, for either the 2D box or the extruded box.

I don't think OCC necessarily expects fragmented inputs, but it does look like there is a behavior difference between fragmented and not for fully-contained 2D entities for intersect.

Some more checks: Case A give no intersection with either volume with metal at the interface, but perturbing the metal z-height by 1nm makes intersection successful. The behavior reproduces with just one metal entity (empty intersection for case A, expected result for case B). The behavior reproduces if we intersect with the boundary of vacuum created by the postrender op (:vacuum_bdy, get_boundary, (:vacuum, 3)) (empty intersection for case A, expected result for case B).

So far I don't see any good options to avoid this on the DeviceLayout side. OCC just seems inconsistent here. For example, the intersection of :vacuum and :substrate is empty both before and after fragmenting. So do volumes "include" their boundary or not? (I think the answer is that the gmsh.model.occ.intersect operation is actually called "Common", and is documented to be empty if the outputs would have lower dimension than both inputs.) Is that also a footgun we should guard against? And occ.intersect(A, B) seems to be different from A - (A - B) where - is occ.cut even with the 2D-only example when B is fully contained in A, which is surprising. (This one is not documented, as far as I can tell. Maybe it is a bug, either in OCC or the Gmsh wrapper?)

We could at least add some kind of warning, but I'm not sure what to say or where (maybe just in the docs for intersect_geom!).