Inference adapter for OpenAI-compatible APIs.
Handles wire protocol translation only:
- Request serialization (messages, tools, provider routing)
- System role demotion at serialization time (profile-driven)
- Response parsing (faithful field mapping, no model-specific fallbacks)
- Retry on transient HTTP errors (429, 502, 503, 529)
Model-specific normalization (content fallback from thinking, XML tool-call
extraction) is handled by Response.Normalizer in the Gateway layer.