an issue is that with this pointers can't be freed with free anymore, so technically an ABI break.
Its probably unlikely but there might be users out there who try that, or pass in malloc pointers to array constructors and then transfer ownership to the array.

I noticed that PyDataMem_RENEW() is obviously wrong the base pointer changes alignment. Fixing, but this adds some overhead in cases where the base offset changes.
(also, there seem to be some problems in 32-bit mode)

This PR and #5470 needs another look from reviewers, @seberg we probably need a "Awaits review" tag for such PRs ?

There is a "ready for review" tag, not sure if it helps here... I guess the problem is getting someone to review who knows this better.
Not sure I am too worried about transferred ownership of allocated pointers to consider it an ABI break). I suppose if we go there, that also opens us up to simply using the Python allocator for the strides+shape (dims), or even moving it into the object allocation to optimize it away entirely?

Anyway, you are right, this is one that could be reactivate, but may need prodding reviewers a bit :).

okay thanks for the summary @seberg. I guess I will tag this and the other pr Ready for Review and revisit this later.

Thanks for this! I will close it, since it is now superseded by gh-17582 and NEP 49 — Data allocation strategies.