test: protect large HTTP payload compatibility (#3721)

## Summary

Protect the current and recent releases’ support for large valid HTTP
responses across Responses and Chat Completions. The payload guidance
follows
[openai-node#2433](https://github.com/openai/openai-node/pull/2433):
supported API output should not gain new arbitrary body, event, or line
rejection limits as a side effect of security or performance fixes.

Add a deterministic public-entrypoint contract test with synthetic 32
MiB+1 text. It covers blocking Responses `output_text`, structured JSON
parsing in both APIs, streamed Responses deltas and final responses, and
Chat Completions' partial-JSON accumulation and final parsed completion.
The cases run sequentially in one test so pytest workers do not multiply
their memory cost. The intentional payload size is a regression probe,
not an API maximum; fixtures stay in memory and use no live API or real
credentials.

The existing HTTP paths accept these payloads, so this PR changes only
agent guidance and a standalone repository-owned test. It makes no
runtime, dependency, generation, or custom-code budget changes.
WebSockets are outside this PR's scope.