Python: [Feature]: Add max_retries (or OpenAI client kwargs) to FoundryChatClient #9719

Description

FoundryChatClient / RawFoundryChatClient obtain their async OpenAI client via project_client.get_openai_client(**openai_kwargs) but only forward default_headers (and an internal feature-usage http_client when the framework constructs the project client). There is no supported way to set OpenAI SDK options such as max_retries at construction time.

In agent-framework-foundry 1.15.0, when a caller passes an existing AIProjectClient via project_client=, the OpenAI client inherits the SDK default max_retries (typically 2). Hosts that share one AIProjectClient across several FoundryChatClient instances and need fail-fast behavior on throttling (no automatic HTTP retries) currently subclass AIProjectClient and override get_openai_client to inject max_retries=0. That works but depends on the internal call path in FoundryChatClient.__init__ continuing to call get_openai_client without retry arguments.

Related: #9698 added timeout on the same code path; max_retries is the same class of gap.

Proposal

Add one of:

1. max_retries: int | None = None on RawFoundryChatClient / FoundryChatClient, forwarded into project_client.get_openai_client(max_retries=..., **openai_kwargs) when set; or
2. A broader openai_client_kwargs: Mapping[str, Any] | None (or **openai_client_kwargs) forwarded to get_openai_client; or
3. Documented support for passing a pre-built async OpenAI client (if that is the preferred extension point).

When not provided, behavior should remain unchanged (SDK defaults).

Language

Python

Context

Tracked internally as CFX-15650 (Cofoundrs Hosted Agent).