memory regression between 0.0.1a5 and 0.0.1a6?

[carljm]

We've had a report of a significant memory regression between these two versions (like from 52GB on a5 to >80GB on a6) on a non-public codebase.

Not obvious in the a6 changelog which PR would be the culprit; may need to see if we can reproduce a memory regression on a smaller codebase and then bisect on that.

[carljm]

This was bisected to astral-sh/ruff#18156

[carljm]

We've been working on memory usage and have brought it down quite a bit, though it's still higher than we'd like. I think the conclusion here was that as best we could tell there wasn't a specific bug introduced in this PR, just more understanding of more types leading to higher memory usage. So I don't think keeping this specific issue open is useful anymore.