[ilan-gold]

This is a vibe-coded solution to #44 (comment) - instead of writing directly into the shared output buffer within the async tasks (which AFAICT is impossible), we just use double the memory. This is what zarrs does anyway internally. I will hopefully have time to clean this up at some point. Definitely open to high-level points, but not sure there is much point in code review now. This is more of a PoC

[ilan-gold]

Investigating this more, we are making a quite fundamental tradeoff here beyond that already enumerated - by writing to the output buffer in rust, we are not allowing the python to call the rust asynchronously (basically smashing up the stack the lifetime problem that motivated the doubling of memory in the first place).

[flying-sheep]

One interesting thing I learned about safety is that numpy doesn’t actually make guarantees against data races – so any memory shared with numpy can not be converted into a rust reference (read or write), but can be written through with raw pointers (* or NonNull or so).