Significantly optimise checking of repeatedly-instantiated TypedDict objects

[alexander-beedie]

Summary

Identified a somewhat niche hotspot relating to TypedDict; with this PR in place, type-checking pydantic gets ~2x faster (measured ~60% less CPU time), and two other packages get a ~6-10% improvement. Pretty much a no-op otherwise.

Problem

Instantiating TypedDict fields rebuilt a full SmallMap of field types on every call, with each rebuild deep-copying every field type; callers that only iterate paid that cost repeatedly.

Solution

• Cache the instantiated field map per Solver, in such a way that repeat calls share one allocation (instead of re-cloning each field's type tree).
• Only cache what is context-independent (mirrors the existing protocol_cache rule of caching only Var-free types).
• Never cache an incomplete map (field lookup failures are now classified: see FieldOutcome).
• Fixed InitVar on a TypedDict member (it's not valid and blocked caching).
• Don't clone ClassField just to read it.

Test Plan

All existing unit tests pass without modification.
Several new unit tests added.

Benchmarks¹

Huge positive impact on pydantic (the pydantic_core.core_schema module is an absolute wall of TypedDict) 🚀

The speedup tracks with repeated instantiation of the same TypedDict, not the overall TypedDict count, and it seems that's not a common pattern; only a couple of packages show much benefit (aside from pydantic the only other solid gains I spotted are for numpy and narwhals, which both get a bit faster).

CPU time (user+sys)

My machine was a bit contended, so I recorded CPU-time instead of wall-clock as it was much more stable (measurements are standard multithreaded runs):

project	before	after	Δ
pydantic	3.71s	1.53s	−58.7%
numpy	12.01s	11.08s	−6.9%
narwhals2	2.36s	2.18s	−7.9%

Footnotes

1. Test machine: Apple Silicon M3 Max (16 cores). ↩

2. @MarcoGorelli, "you're welcome" 🤣 ↩

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D115363524. (Because this pull request was imported automatically, there will not be any future comments.)

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[yangdanny97]

@samwgoldman

[yangdanny97]

TypedDict stuff would normally be under my wheelhouse but I don't have any capacity in the near future so I'll let Sam handle this