BUG: same_kind casting between strings and bytes is inconsistent #32880

Describe the issue:

I would have expected that the different unicode strings could be casted to each other with "same_kind", but not to and from bytes (since really they are not the same kind). Oddly, however, casting strings to bytes is always possible, but the reverse is not possible for U to S.

Found while commenting on #32757

Reproduce the code example:

import numpy as np
s = np.array(b"a", dtype="S")
u = np.array("a", dtype="U")
t = np.array("a", dtype="T")

# U->S
np.can_cast(u.dtype, s.dtype, casting="same_kind")
# False  # Seems logical
u.astype(s.dtype, casting="same_kind")
# TypeError: Cannot cast scalar from dtype('<U1') to dtype('S1') according to the rule 'same_kind'
# OK consistent
# S->U
np.can_cast(s.dtype, u.dtype, casting="same_kind")
# True  # Huh, not the same?
s.astype(u.dtype, casting="same_kind")
# array('a', dtype='<U1')
# Why does this work?

# T->S
np.can_cast(t.dtype, s.dtype, casting="same_kind")
# True  # Huh? Why different from U->S?
t.astype(s.dtype, casting="same_kind")
# array(b'a', dtype='|S1')
# S->T
np.can_cast(s.dtype, t.dtype, casting="same_kind")
# True
s.astype(t.dtype, casting="same_kind")
# array('a', dtype=StringDType())

Error message:

Python and NumPy Versions:

2.6.0.dev0+git20261005.414e69e
3.14.7 (main, Sep 13 2026, 08:30:29) [GCC 16.2.0]

Runtime Environment:

• numpy_version: 2.6.0.dev0+git20261005.414e69e
  python: 3.14.7 (main, Sep 13 2026, 08:30:29) [GCC 16.2.0]
  uname:
  machine: x86_64
  node: dolphin
  processor: ''
  release: 7.1.13+deb14-amd64
  system: Linux
  version: 'Misc #1 SMP PREEMPT_DYNAMIC Debian 7.1.13-1 (2026-09-03)'
• simd_extensions:
  baseline:
  • X86_V2
    found:
  • X86_V3
    not_found:
  • X86_V4
  • AVX512_ICL
  • AVX512_SPR
• ignore_floating_point_errors_in_matmul: false

How does this issue affect you or how did you find it:

No response