Fix audit findings A-H: ctypes boundary, DLR basis functions, and test-suite self-sabotage

[shinaoka]

Re-verified each of the eight audit findings against mainline (v2.1.3) before
fixing it, per AGENTS.md -> spm-agent-rules (rules/ffi-boundary.md,
rules/python.md, rules/testing.md).

Disposition

#	Finding	Disposition
A	DLR .u/.uhat returned the IR basis functions, not the DLR ones	Fixed properly. Re-verified: spir_basis_get_u/spir_basis_get_uhat accept the DLR spir_basis* handle, so no NotImplementedError and no C-API issue is needed. Both properties now build the DLR function sets lazily from the DLR handle, and g_dlr @ dlr.u(tau) reproduces gl @ basis.u(tau) (asserted in the new tests).
B	No dtype-width normalization at the boundary	Fixed. Swept every .ctypes.data_as site in the package (sampling.py, dlr.py, poly.py; basis.py, sve.py, kernel.py, augment.py have none). Each input goes through a new _util.as_boundary_real/as_boundary_complex/as_boundary_matsubara helper (np.ascontiguousarray(x, dtype=<explicit>)) with the pointer taken from the converted object.
C	DLR __init__ pointer provenance	Fixed. Poles are normalized first and the pointer is taken from the normalized array; shape/emptiness/finiteness are validated.
D	FunctionSet.__call__ squeeze bugs	Fixed. Documented contract: the result is (n_funcs,) + np.shape(x), the leading axis is dropped only for a single function, the point axes only for scalar x. FunctionSetFT.__call__ follows the same rule. This is what the previously-skipped test_zero_frequency expected (basis.v(np.array([0.0])).shape == (size, 1)).
E	rescale raised NameError (undefined new_lambda)	Fixed. Holds lambda_ == beta * wmax fixed, computes new_wmax = lambda_ / new_beta, reuses the existing kernel and SVE (both depend only on lambda_ and eps), and rejects non-positive new_beta with ValueError.
F	Fermionic TauConst gives a rank-deficient basis	Raises now. ValueError naming the reason: hat(n) == sqrt(beta) * (n == 0) and fermionic reduced frequencies are odd, so the augmentation column is identically zero. The two fermionic parametrizations in test_augment.py were converted into a pytest.raises test.
G	Matsubara index truncation and i % n index wrap-around	Fixed. check_reduced_matsubara rejects complex input (TypeError), non-finite and non-integral values (ValueError naming the offending value, e.g. 1.9), and wrong parity (naming "odd"/"even"). FunctionSetFT.__call__ validates before converting to int64. Both __getitem__ implementations now use _util.resolve_function_indices: negative indices resolve with Python semantics, out-of-range raises IndexError, non-integral raises ValueError.
H	Test-suite self-sabotage	Fixed. See below.

Additional tracking issues opened for things deliberately not forced in this PR:
#82 (should TauLinear also be bosonic-only?), #83 (restore the
SVEResult.part()-based coverage deleted from test_poly.py), #84 (basis
truncation basis[:n] is still unimplemented in the C API).

Finding H in detail

• The 8 skips in tests/test_sve_advanced.py were not legitimate
  preconditions. They were except Exception: pytest.skip(...) blocks hiding a
  bug in the tests themselves: sve_result_new/basis_new were called with the
  Python LogisticKernel/RegularizedBoseKernel wrapper instead of
  kernel._ptr (TypeError: expected LP_LP__spir_kernel instance instead of LogisticKernel), and basis_new was called with 5 of its 7 arguments. Fixed
  and un-skipped; all 8 now run and pass.
• Removed the try/except -> pytest.skip wrappers in
  tests/test_advanced_features.py (10 of them), tests/test_sampling_advanced.py,
  and the except Exception: print(...) in the test_bases fixture in
  tests/conftest.py, which handed tests a silently incomplete dict.
• tests/test_poly.py: deleted the two dead triple-quoted blocks (8 stringified
  tests depending on the unavailable SVEResult.part(), tracked in Restore SVEResult.part()-based poly coverage once the C API exposes it #83), deleted
  the shadowed duplicate test_broadcast_uv that referenced undefined
  beta/atol, and added the missing assert on the no-op comparison at line 22.
• The suite now has no skips at all.

Regression tests

New tests/test_ffi_boundary.py (68 tests), plus the converted augmentation tests:

• dtype matrix — float32/float64/complex64/complex128 and integer
  input for TauSampling.evaluate/fit, MatsubaraSampling.evaluate/fit,
  DLR.from_IR/to_IR, and FunctionSet.__call__; each compared against the
  float64 reference at the input dtype's own precision (relative to the norm,
  since the coefficient vectors span many decades).
• non-contiguous input for every one of those entry points, and for DLR poles.
• concrete exception types with match= — non-integral Matsubara index
  (1.9), wrong parity, complex indices, complex evaluation points, NaN input,
  out-of-range axis, out-of-range and non-integral function index, empty /
  multi-dimensional sampling points, non-positive rescale beta, fermionic
  TauConst.
• nonzero-norm assertions throughout, so a silently zeroed result cannot pass.
• shape tests for the finding-D semantics (whole set, single function, slice,
  scalar and tensorial x, for both u and uhat).
• finding A correctness: g_dlr @ dlr.u(tau) and g_dlr @ dlr.uhat(n)
  checked against the IR result.
• axis coverage (0, 1, -1, -2) with round-trip checks.
• rescale round-trip (lambda_ preserved, spectrum preserved, basis usable).
• __all__-derived smoke test and a from sparse_ir import * test.

Test counts: 101 passed / 8 skipped -> 181 passed / 0 skipped.

Other defects fixed in passing

Negative axes are resolved before reaching C (which takes a non-negative target
dimension) and out-of-range axes raise IndexError; non-finite input is rejected
before it reaches a factorization; returned handles are null-checked and statuses
compared against COMPUTATION_SUCCESS; the bare except: pass in both
release() methods and a stray print(type(...)) in PiecewiseLegendrePoly.overlap
are gone; basis._slice_to_size (an AttributeError at both call sites in
augment.py) is replaced by a local range-checking _slice_to_size;
AugmentedTauFunction.xmax returned xmin; AugmentedTauFunction.deriv passed
*daug where the constructor takes a list; and __all__ no longer names the
nonexistent TauPoles/MatsubaraPoles, which made from sparse_ir import *
raise AttributeError.

Behavior changes worth reviewer attention

• TauSampling on a MatsubaraConst-augmented (vertex) basis now raises
  ValueError explaining that the augmentation is undefined in imaginary time,
  instead of handing NaNs to the C factorization. No existing test covered this
  path; a new one does.
• TauConst(beta, 'F') now raises instead of building a rank-deficient basis.
• Non-integral or wrong-parity Matsubara indices now raise instead of being
  truncated or accepted.

No REPOSITORY_RULES.md override applies to this change beyond what it already
documents (status codes from pylibsparseir.constants, no local CDLL, and the
version-consistency check).

Version

pyproject.toml 2.1.3 -> 2.1.4; python check_libsparseir_version_consistency.py
passes (pylibsparseir>=0.8.3,<0.10.0 in both pyproject.toml and
.conda/meta.yaml).

🤖 Generated with Claude Code