refactor[next]: carry the connectivity name in OffsetType

[havogt]

Stacked on #2724 — the first commit here is #2724's. Please review/merge that one first; this PR's own change is the second commit (refactor[next]: carry the connectivity name in OffsetType).

Description

Follow-up to #2724, which patched the symptom. This removes the cause.

The unstructured shift lowering took the ITIR offset tag from the source-level identifier (foast.Name.id, and after #2724 also foast.Attribute.attr), so field(Off) had to pattern-match the FOAST node shape, and resolved to the wrong tag whenever the identifier was not the connectivity name:

from next_tests.integration_tests.cases import V2E as RenamedV2E

@gtx.field_operator
def testee(a: cases.EField) -> cases.VField:
    return neighbor_sum(a(RenamedV2E), axis=V2EDim)

embedded: [1 3]
gtfn:     KeyError: "Offset 'RenamedV2E' not found in offset provider."

Embedded resolves through FieldOffset.value (fbuiltins.py:485-488) and was unaffected, so the two execution paths disagreed. This dates back to the first FOAST shift lowering (#625, 2022-01-26) and became observable when embedded remap landed (#1309).

Change

ts.OffsetType gains name, set from FieldOffset.value in __gt_type__. Both shift arms in _visit_shift now match on the type rather than the node shape — mirroring the Cartesian arm, which already reads the Dimension out of ts.DimensionType. The foast.Attribute special case from #2724 is dropped; it is subsumed.

name is None for the Dim + idx offsets synthesized in _deduce_binop_type. Those are Cartesian, resolved from source/target alone, and never looked up in the offset provider. FieldOffset.__gt_type__ always sets it, so Cartesian FieldOffsets keep a name and the field(Off[idx]) arm is unaffected.

The field(Off) arm moved below the as_offset arm: as_offset(...) propagates its argument's OffsetType, so it would otherwise be caught by the now type-driven pattern. (It also has a single target, so the target=[_, _] guard would reject it — the move makes the ordering intentional rather than incidental.)

This is a step toward the # TODO(havogt): replace by ConnectivityType already on OffsetType, not the whole thing — I did not add an ADR since it does not change the public API shape.

Not addressed

A separate name coupling is still live in the backends: a sparse field's LOCAL dimension is looked up in the offset provider by its value, so the local dimension must be named exactly like its offset.

• codegens/gtfn/gtfn_module.py:93-96 — dim_name = dim.value → common.get_offset_type(offset_provider_type, dim_name)
• runners/dace/lowering/gtir_to_sdfg.py:553 — raises ValueError("The provided local dimension … does not match any offset provider type.")
• runners/dace/lowering/gtir_dataflow.py:733

That is the direction of common.py:977-979 (TODO(havogt): refactor towards encoding this information in the local dimensions of the ConnectivityType.domain) and is untouched here.

Tests

Added to test_import_from_mod.py, across the backend matrix:

• test_import_renamed_offset_unstructured_shift — neighbor_sum(a(RenamedV2E), axis=V2EDim)
• test_import_renamed_offset_sparse_shift — a(RenamedE2V[0]), covering the field(Off[idx]) arm

Both fail on #2724 and pass here.

Local runs (full backend matrix incl. GPU, mypy src/ and pre-commit clean):

• tests/next_tests/unit_tests — 2151 passed, 9 skipped, 15 xfailed
• tests/next_tests/integration_tests — 5316 passed, 197 skipped, 22 deselected, 612 xfailed

The 22 deselected are test_orchestration.py::test_sdfgConvertible_connectivities[*-dace.run_dace_gpu] and its module siblings: that test fails identically with and without this change on my machine (TypeError: Illegal copy! (from gt_conn_E2V to tlet_0_deref)), verified by stashing and re-running on the base commit. Pre-existing, unrelated.

Requirements

• All fixes and/or new features come with corresponding tests.
• Important design decisions have been documented in the appropriate ADR.

[Copilot]

Pull request overview

This PR refactors the gt4py.next frontend type system so unstructured shift lowering uses the connectivity’s declared name carried in ts.OffsetType, rather than relying on the source-level identifier shape (e.g., foast.Name vs foast.Attribute). This resolves mismatches when offsets are module-qualified or renamed on import, aligning compiled lowering behavior with the embedded path.

Changes:

• Add name: Optional[common.Tag] to ts.OffsetType and set it from FieldOffset.value during __gt_type__.
• Update _visit_shift lowering to pattern-match on ts.OffsetType(name=...) (type-driven) for both field(Off) and field(Off[idx]), and adjust match ordering around as_offset(...).
• Extend integration test coverage to include renamed offsets for both unstructured neighbor shifts and sparse shifts across the backend matrix.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/next_tests/integration_tests/feature_tests/ffront_tests/test_import_from_mod.py	Adds regression tests for renamed/module-imported offsets in unstructured and sparse shift lowering.
src/gt4py/next/type_system/type_specifications.py	Extends OffsetType with an optional name to carry the connectivity tag used for offset-provider lookup.
src/gt4py/next/ffront/foast_to_gtir.py	Switches unstructured shift lowering to type-driven matching using OffsetType.name, removing node-shape coupling.
src/gt4py/next/ffront/foast_passes/type_deduction.py	Preserves OffsetType.name when slicing unstructured offsets (e.g., Off[idx]).
src/gt4py/next/ffront/fbuiltins.py	Ensures FieldOffset.__gt_type__ populates OffsetType.name from FieldOffset.value.