fix: don't crash at startup when OPENAI_API_KEY is unset

[lvjr3383]

Problem

InferenceAPI.__init__() eagerly constructs all provider clients at instantiation time, including four OpenAI-specific classes:

File	Line	Failing call
openai/base.py	72	openai.AsyncClient(base_url=...)
openai/embedding.py	23	openai.AsyncClient()
openai/moderation.py	30	openai.AsyncClient()
openai/s2s.py	49	os.environ["OPENAI_API_KEY"]

openai v1+ raises OpenAIError at construction time if OPENAI_API_KEY is not in the environment — even if the caller only intends to use Anthropic, Together, or Gemini models. The result is an immediate crash before any model routing occurs:

openai.OpenAIError: The api_key client option must be set either by
passing api_key to the client or by setting the OPENAI_API_KEY
environment variable

This affects every researcher running the alignment-faking paper replication with Anthropic or Together models and no OpenAI key.

Fix

Replace hard environment access with a sentinel fallback:

# before
self.aclient = openai.AsyncClient()                  # crashes
self.api_key = os.environ["OPENAI_API_KEY"]          # crashes

# after
self.aclient = openai.AsyncClient(
    api_key=os.environ.get("OPENAI_API_KEY", "not-configured")
)
self.api_key = os.environ.get("OPENAI_API_KEY", "not-configured")

The client now initialises without error. If a caller actually invokes an OpenAI model without a real key, the failure happens at call time with HTTP 401 — which is clear and expected.

Testing

Verified on open-source-alignment-faking with no OPENAI_API_KEY set, running claude-3-5-sonnet-20240620. Before this patch: crash at line 1. After: pipeline initialises, loads dataset, and reaches API call execution.