[jason-schrader-openrouter]

Summary

Fixes #1056.

utils/serializers.py calls pydantic.create_model("Unmarshaller" | "Marshaller", body=(typ, ...), __config__=...) every time marshal_json or unmarshal runs. As a result, each chat.send / chat.send_async call builds and compiles about 19 throwaway model classes. That file is Speakeasy template code and gets overwritten on regeneration, so this PR fixes it from _hooks/, which regeneration doesn't overwrite:

• New file _hooks/serializer_cache.py. install() replaces serializers.create_model with a version that builds one model per (name, type, config) and keeps it in a bounded lru_cache (_CACHE_SIZE = 1024).
• The cache key is _type_key(typ), not typ alone. Annotations that compare equal can still validate differently, so the key keeps union member order and the type of each leaf:

  int | float == float | int        # True, but pydantic tries members in order
  Literal[1] == Literal[True]       # True, but pydantic checks the literal's type
  _type_key(int | float)  # (UnionType, ((type, int), (type, float)))

  Without this, unmarshal("1", int | float) followed by unmarshal("1", float | int) returned 1 (int) from the cache instead of 1.0 (float).
• Calls in any other shape, or with arguments that can't be hashed, go straight to pydantic.create_model. If the template changes, the SDK keeps working; it just stops caching.
• init_hooks() in registration.py calls install(), so the cache is active as soon as an OpenRouter client is created.

Measured on 1.3.173 with a local stub server (Python 3.12, pydantic 2.14.3):

	Models built per call	CPU per call
Before	19	15–17 ms
After	0	4–5 ms

tests/test_serializer_cache.py checks that:

• creating a client installs the hook, and a second chat.send builds no new models;
• for unions in both orders, literals, aliases, Optional, generics and validation errors, unmarshal and marshal_json give the same result type, value or error with the cache as with plain pydantic.create_model, after warming the cache in forward and reverse order;
• unhashable annotations fall back without caching;
• the cache stops growing at _CACHE_SIZE.

The equivalence and bound tests fail against the previous cache key. Delete this hook once the upstream template caches these models itself.

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring