Allow injecting embeddings and chat clients

[TheovanKraay]

Summary

Adds optional embeddings_client and chat_client constructor arguments to both
CosmosMemoryClient and AsyncCosmosMemoryClient. When supplied, the toolkit uses
the injected client instead of constructing an Azure-backed one, and treats the
caller as the owner of its lifecycle (so close() does not close it). When omitted,
behavior is unchanged: the toolkit builds and owns the Azure OpenAI / AI Foundry
clients exactly as before.

Motivation

Today the embeddings and chat clients are hard-constructed inside __init__, with no
supported way to supply your own. That couples the toolkit to Azure OpenAI / AI Foundry
and forces anyone who needs a different backend (or offline/deterministic behavior) to
reach into private internals (_embeddings_client / _chat_client). This adds a clean,
public seam instead.

Use cases this enables:

• Pluggable model backends — OpenAI-compatible or self-hosted embedding/chat
  services (via an OpenAI-compatible client), gateways, or alternate providers.
• Reuse of a caller-configured client — apps that already build a client with
  custom retry/timeout/proxy/telemetry can hand the same instance to the toolkit
  instead of it creating a second, differently-configured one.
• Deterministic, offline testing — inject fakes to exercise the pipeline and the
  Cosmos vector path (e.g. against the Cosmos DB emulator) without any Azure account,
  and without monkeypatching private internals.

Changes

• CosmosMemoryClient.__init__ and AsyncCosmosMemoryClient.__init__ accept
  embeddings_client and chat_client (both default None).
• Ownership tracked via _owns_embeddings_client / _owns_chat_client, mirroring the
  existing _owns_*_credential pattern. close() only closes clients the toolkit built.
• Unit tests (sync + async) covering: injected clients are used, ownership flags,
  independent injection of one client, and that close() does not close injected clients.
• CHANGELOG entry under Unreleased.

Compatibility

Fully backward compatible. Both arguments default to None; existing callers get the
current Azure-backed clients and lifecycle with no changes.

Testing

• tests/unit/test_cosmos_memory_client.py and tests/unit/aio/test_cosmos_memory_client.py
  pass (128 tests across both suites).
• ruff check and ruff format --check clean on the changed files.

[Copilot]

Pull request overview

Adds a supported dependency-injection seam for the model-layer clients used by the Cosmos memory clients, allowing callers to supply their own embeddings/chat backends (or fakes) while preserving existing default Azure-backed behavior and lifecycle semantics.

Changes:

• Add optional embeddings_client and chat_client constructor args to CosmosMemoryClient and AsyncCosmosMemoryClient, with ownership flags controlling whether close() tears them down.
• Update close() implementations to only close clients constructed by the toolkit.
• Add sync/async unit tests and a CHANGELOG entry documenting the new injection capability.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
azure/cosmos/agent_memory/cosmos_memory_client.py	Adds injectable embeddings/chat clients and ownership-aware closing for the sync client.
azure/cosmos/agent_memory/aio/cosmos_memory_client.py	Adds injectable embeddings/chat clients and ownership-aware closing for the async client.
tests/unit/test_cosmos_memory_client.py	Adds sync unit tests verifying injection behavior and lifecycle ownership.
tests/unit/aio/test_cosmos_memory_client.py	Adds async unit tests verifying injection behavior and lifecycle ownership.
CHANGELOG.md	Documents the new injection feature under Unreleased.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.