[AryanRG]

NumPy clip raises ValueError: assignment destination is read-only when an integer array is clipped with same-dtype bounds from broadcast_to or an array marked read-only.

Skip the range adjustment when a bound already has the input dtype, since its values are representable. Regression tests cover lower and upper bounds, broadcasted bounds, and out= without changing the bound or input.

Validated the NumPy tests on NumPy 1.26.4 and 2.5.4, plus the local suite with NumPy and Torch. Optional backends were skipped.

[ev-br]

CI failures:

• test_matmul under numpy 1.22 is unrelated and not worth spending time on, TST: skip test_matmul on NumPy 1.22 because @= fails #487 adds a skip.
• dpnp testing failures are also unrelated. pypi dpnp page shows activity on Oct 28, I wonder if something changed in dpnp itself. Any insights @Nin17 @antonwolfy?

[ev-br]

Apart from unrelated failures, the patch LGTM.

[ev-br]

Merged, thank you @AryanRG !

[antonwolfy]

@ev-br

This comes from dpnp 0.21.0, released on Oct 28. It moved to the 2025.12 version.
dpnp's __array_namespace__ accepts only the version it implements, so api_version="2024.12" now raises ValueError: Only 2025.12 is supported.

I believe this is intended on the dpnp side. And so the fix is to add "2024.12" to the dpnp xfail list in tests/test_array_namespace.py. I'm happy to open a PR for it.

[ev-br]

Happy to review a PR for skipping 2024.12!