[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, typ, config) and caches it with lru_cache.
• 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

Responses were identical before and after.

tests/test_serializer_cache.py checks that:

• creating a client installs the hook;
• a second chat.send builds no new models;
• cached models serialize the same way as uncached ones.

The first two tests fail if the install() call is removed. 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