fix: bound TypeAdapter lru_cache to prevent memory leak in multi-threaded usage

[giulio-leone]

Summary

Fixes #2672

Problem

_CachedTypeAdapter uses lru_cache(maxsize=None) (unbounded), which causes memory leak in multi-threaded environments. When responses.parse is called from multiple threads, generic types like ParsedResponseOutputMessage[MyClass] are created anew by pydantic in each thread, generating distinct cache keys. This means the cache never hits and grows without bound.

Root Cause

In src/openai/_models.py line 802:

_CachedTypeAdapter = cast('TypeAdapter[object]', lru_cache(maxsize=None)(_TypeAdapter))

Python's typing module creates new generic type objects in different threads, so ParsedResponseOutputMessage[Fact] in thread A has a different identity than ParsedResponseOutputMessage[Fact] in thread B. Since lru_cache keys on object identity, each thread creates a new cache entry.

Fix

Set maxsize=4096 to cap memory usage while still providing effective caching for typical workloads. This is a standard practice for lru_cache — the LRU eviction policy ensures the most-used types stay cached while preventing unbounded growth.

Testing

Verified that the cache is bounded:

info = _CachedTypeAdapter.cache_info()
assert info.maxsize == 4096

[giulio-leone]

Friendly ping — CI is green and this is ready for review. Happy to address any feedback. Thanks!