tsp_asymmetric segfaults (SIGSEGV) on a basic dense cost matrix

[3x3xX3N0N]

tsp_asymmetric crashes the interpreter with a segmentation fault (exit 139) on a plain dense asymmetric cost matrix. It reproduces in a base install (no pyopencl), so it isn't GPU/OpenCL-related.

Repro (segfaults verbatim)

import numpy as np
from pyqrackising import tsp_asymmetric
M = np.random.default_rng(3).random((200, 200)).astype(np.float32)
np.fill_diagonal(M, 0.0)
print(tsp_asymmetric(M, quality=3, k_neighbors=20))

Output: Segmentation fault, process exits 139, no Python traceback.

Expected

Returns a (tour, length) for the asymmetric instance, as tsp_symmetric does for symmetric ones.

Ruled out (all still crash)

• Not OpenCL/GPU — crashes with pyopencl uninstalled (CPU path) and installed.
• Not dtype — crashes with float32 and float64.
• Not multi_start — crashes with multi_start=1 and 4.
• Not size — crashes at n=50 and n=200.
• Specific to tsp_asymmetric — tsp_symmetric and maxcut_tfim run fine on the same machine and instances.

Environment

• PyQrackIsing 9.16.2
• Python 3.12.10, Windows 11 (10.0.26200)
• numpy 2.4.6, numba 0.66.0, scipy 1.18.0, networkx 3.6.1

Not yet tested on Linux, so it may be platform- or numba-version-specific. Happy to run a faulthandler / debug trace if that would help localize it.

[3x3xX3N0N]

Update — does NOT reproduce on Linux; appears Windows-specific

Ran the identical repro on Linux with the same package versions (numba 0.66.0, numpy 2.4.6):

OS	Python	Result
Windows 11 (10.0.26200)	3.12.10	Segfault (exit 139)
Linux — WSL2 Ubuntu	3.14.4	✅ works, returns a valid tour
Linux — Docker python:3.12-slim	3.12	✅ works (exit 0)

So it's Windows-specific, not Python-version-related (3.12 is fine on Linux). Same tsp_asymmetric(M, quality=3, k_neighbors=20) call throughout. This points at a Windows-only native / numba codegen or memory issue in the asymmetric path rather than a logic bug. Happy to attach a faulthandler trace from the Windows crash if that would help localize it.