Fix real-input Matsubara fit and complex DLR at the ctypes boundary

[shinaoka]

Summary

Fixes two ctypes-boundary defects that make real-input Matsubara fitting and complex-coefficient DLR unusable, plus a patch version bump.

1. MatsubaraSampling.fit with real-dtype input produced silent garbage

src/sparse_ir/sampling.py, MatsubaraSampling.fit (around line 316) passed the input buffer straight to spir_sampling_fit_zz (the complex128 C entry point) without normalizing dtype first. A float64 array was reinterpreted as complex128 — i.e. its raw byte buffer was read as though it were twice as long as actually allocated. This is an out-of-bounds read that does not raise: it silently returns finite-looking-but-wrong (in practice nan) values.

Failure mode: silent numerical garbage, no exception.

Fix: normalize ax to complex128 via np.ascontiguousarray(ax, dtype=np.complex128) before taking the pointer, matching the pattern already used by TauSampling.fit/evaluate and MatsubaraSampling.evaluate. Added a dtype guard (ax.dtype.kind not in ("f", "c")) that raises ValueError for unsupported dtypes instead of silently misreading them.

2. Complex DLR (from_IR/to_IR) raised a ctypes.ArgumentError

src/sparse_ir/dlr.py, DiscreteLehmannRepresentation.from_IR/to_IR (lines ~137–148 and ~194–205) cast complex128 buffers as ctypes.c_double pointers (with a leftover # TODO: use complex data) when calling spir_ir2dlr_zz/spir_dlr2ir_zz, which actually expect c_double_complex pointers.

Failure mode: ctypes.ArgumentError: argument 7: TypeError: expected LP_c_double_complex instance instead of LP_c_double — complex DLR coefficients were completely unusable.

Fix: mirror the pattern used in sampling.py: convert input to complex128 contiguous arrays, allocate the output as a c_double_complex-dtyped structured array, pass proper POINTER(c_double_complex) pointers to the _zz entry points, and reassemble the output as output['real'] + 1j * output['imag']. Confirmed pylibsparseir exposes the needed complex-capable bindings (spir_ir2dlr_zz, spir_dlr2ir_zz) with LP_c_double_complex argtypes.

Regression tests

• tests/test_sampling.py: added test_evaluate_fit_roundtrip_complex to both TestTauSampling and TestMatsubaraSampling (rng seed 42, asserts np.iscomplexobj and max error < 1e-12 including .imag).
• tests/test_sampling.py: added TestMatsubaraSampling::test_fit_real_input_matches_complex_reference, the regression test for defect 1 — fits a real float64 array and checks it agrees with fitting the same values cast to complex128 (pre-fix this produced nan).
• tests/test_dlr.py: added test_complex_roundtrip, the regression test for defect 2 — builds complex IR coefficients from known poles (mirroring test_compression's construction so the round trip is expected to be near-exact) and checks from_IR/to_IR round-trips within 300*eps (pre-fix this raised ctypes.ArgumentError).

Verified both defects experimentally against the pre-fix code: defect 1's test failed with nan values (silent garbage), and defect 2's test failed with ctypes.ArgumentError (TypeError), exactly matching the reported failure modes.

Version bump

pyproject.toml: 2.1.2 → 2.1.3 (patch bump for bugfix release). .conda/meta.yaml reads its version dynamically from pyproject.toml via load_file_regex, so no separate edit was needed there. check_libsparseir_version_consistency.py passes.

Test results

• Before fix (new regression tests only, run against unfixed code): 2 failed (test_fit_real_input_matches_complex_reference, test_complex_roundtrip), matching the described failure modes.
• After fix: full suite 101 passed, 8 skipped (pre-existing skips in test_sve_advanced.py, unrelated to this change).

🤖 Generated with Claude Code