Fix constant integer overflow in irk_rand_uint32_vec

Declare the shift variable as npy_uint32 instead of npy_int32. The
right-hand side ((npy_uint32)INT_MAX + 1) equals 2**31, which does not
fit in a signed 32-bit integer and wrapped to INT32_MIN when stored.

All downstream uses (lo - shift, hi - shift + 1U, res[i] += shift)
already operate on the unsigned bit pattern via modulo-2**32 arithmetic,
so behavior is bit-for-bit identical. This removes the Coverity
INTEGER_OVERFLOW finding (CID 652701) and the out-of-range signed
conversion it relied on.

Author: antonwolfy

CHANGELOG.md

@@ -14,6 +14,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 * Fixed compatibility with NumPy 2.5 by replacing the deprecated in-place array `shape` assignment with `reshape`, and by replacing the deprecated `numpy.testing.suppress_warnings` usage in tests with `pytest.warns` [gh-137](https://github.com/IntelPython/mkl_random/pull/137)
+* Fixed a constant integer overflow in `irk_rand_uint32_vec` by declaring the `shift` variable as `npy_uint32` instead of `npy_int32`, so `2**31` no longer overflows a signed 32-bit integer (behavior is unchanged as all downstream uses rely on modulo-2^32 arithmetic)
 
 ## [1.4.1] (05/11/2026)
 

mkl_random/src/mkl_distributions.cpp

@@ -1981,7 +1981,7 @@ void irk_rand_uint32_vec(irk_state *state,
 
     if (hi >= intm) {
 
-        npy_int32 shift = ((npy_uint32)intm) + ((npy_uint32)1);
+        npy_uint32 shift = ((npy_uint32)intm) + ((npy_uint32)1);
         int i;
 
         /* if lo is non-zero, shift one more to accommodate possibility of hi