I proposed a different fix in PR #PR 26359. I didn't notice that you was working on a fix in parallel :-)

Why not keep the assertion but add a null check for the return value of winner->tp_new and return -1 from type_new_get_bases?

Then you keep the invariant of "-1 means error, 0 means didn't assign to *type, 1 means assigned to *type"

I proposed a different fix in PR #PR 26359. I didn't notice that you was working on a fix in parallel :-)

No worries. This is funny and slightly sad at the same time but I'm glad either way some kind of fix is going in.

Why not keep the assertion but add a null check for the return value of winner->tp_new and return -1 from type_new_get_bases?

Then you keep the invariant of "-1 means error, 0 means didn't assign to *type, 1 means assigned to *type"

Oops our messages were posted at the same time and I missed yours. Sorry.

@sweeneyde I considered that. But I chose to delete the assertion because that's the old behavior before the refactor (which caused this regression), and I didn't want to take any chances: ecf14e6#diff-1decebeef15f4e0b0ce106c665751ec55068d4d1d1825847925ad4f528b5b872L2464
I felt it was safer to restore the old behavior.

Anyways, Victor's GH-26359 does what you're suggesting. So you can consider his PR instead.

Victor's PR has been merged, so I'm closing this