feat(llm): support OpenAI Decisions evaluation

[wolfy-j]

Applications can evaluate predicates, fixed choices, and ordinal scores through OpenAI's Decisions endpoint using the existing llm.evaluate(state, questions, options) interface. The native OpenAI driver now implements the evaluator contract and reuses its credentials, HTTP transport, retry policy, and error classification.

The adapter sends text or JSON evidence to /v1/decisions, uses opaque question names, and maps answers back to caller slots. It converts OpenAI's zero-based scores to the existing one-based reading contract while preserving full distributions and separate confidence. Refusals fail the whole evaluation; incomplete or malformed answers cannot return successful partial readings. Unsupported generation options fail explicitly. Usage counters retain disjoint input/cache categories and raw provider usage for accounting. A small client guard converts scalar/null JSON responses into errors before metadata attachment.

The README documents direct calls, model discovery, predicate criteria, score meaning, refusals, supported inputs/options, endpoint-specific pricing, and the dedicated live-test flag. This integration supports text/JSON evidence; typed image inputs require a future contract extension. The official API is currently beta with gpt-6-luna support. References: guide, API reference.

Validation: full LLM lint passes on signature-verified Wippy 0.3.45a (125 entries), module manifest checks pass, package dry run passes, and Astra source review found no code blockers. All CI checks pass, including the full 1,632-test LLM suite (run). New tests cover mapping and malformed/refused responses, usage consistency, HTTP retries, facade/model routing, and native versus compatibility contract binding. Live API tests are opt-in and have not been run; no paid API calls were made.