feat(eval): Azure OpenAI support for reader/judge clients (#19)

[jphein]

Summary

• _default_client() in answer_generator.py and longmemeval_judge.py now detects AZURE_API_KEY + AZURE_API_BASE env vars and constructs AzureOpenAI instead of OpenAI. Falls back to existing OPENAI_API_KEY path when Azure vars aren't set.
• run_longmemeval_mempalace.py script defaults flipped to Azure-deployed model names (gpt-4o-mini reader, gpt-4o judge). Library-level defaults unchanged for direct-OpenAI users.
• 8 new Azure-path tests + fixes for 2 pre-existing tests that broke when AZURE_API_KEY is in the shell environment.

Dry-run cost estimate (500 questions)

Component	Model	Input tokens	Output tokens	Cost
Reader	gpt-4o-mini	~761K	~25K	$0.13
Judge	gpt-4o	~150K	~10K	$0.47
Total				$0.60

Live 500-question run not included — ready to fire on go-ahead.

Test plan

• 317 tests pass (8 new Azure-path tests)
• Dry-run cost estimate against longmemeval_oracle.json
• Live daemon run (pending go-ahead — ~$0.60)

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces Azure OpenAI support by updating the client initialization logic in answer_generator.py and longmemeval_judge.py to check for Azure-specific environment variables. It also updates the default models in run_longmemeval_mempalace.py to Azure-compatible versions and adds corresponding unit tests. Feedback focuses on improving error handling for partially configured environment variables to ensure the application fails fast, and refactoring duplicated client initialization and test logic into shared utilities to reduce maintenance overhead.

[gemini-code-assist]

According to the general rules, when fetching credentials from environment variables, we should explicitly check for their presence and raise a RuntimeError if they are missing to ensure the application fails fast. The current implementation silently falls back to the OpenAI path if Azure environment variables are only partially set, which can lead to confusing behavior.

    azure_key = os.environ.get("AZURE_API_KEY")
    azure_base = os.environ.get("AZURE_API_BASE")
    if (azure_key is not None) ^ (azure_base is not None):
        raise RuntimeError("Both AZURE_API_KEY and AZURE_API_BASE must be set for Azure OpenAI support.")
    if azure_key and azure_base:

References

1. When fetching credentials from environment variables, explicitly check for their presence and raise a RuntimeError if missing to ensure the application fails fast.

[gemini-code-assist]

This _default_client implementation is identical to the one added in sme/eval/answer_generator.py. Duplicating complex logic like this increases maintenance overhead and the risk of divergence. Consider refactoring this into a shared utility function within the sme.eval package.

[gemini-code-assist]

The Azure client tests and the _install_fake_openai_module helper are duplicated from tests/test_answer_generator.py. Consolidating the client factory logic into a shared utility would also allow for consolidating these tests, ensuring consistent test coverage across the repository.