feat(m7.5c): TypeScript OpenAI-compatible model adapter

[shamspias]

The TS package had only FakeModel; this is its first real LLM adapter — and since the global fetch is built in, it lives in the core package with ZERO new deps.

• openai.ts: OpenAICompatModel over the global fetch, speaking the Chat Completions dialect — one adapter for OpenAI, Groq, Gemini (compat), Z.ai/GLM, Ollama, and vLLM. Sends the model only each capability's intent surface (ToolSpec) — never policy metadata (§2.5); maps text + tool_calls responses (tool calls win over a 'stop' finish); malformed tool JSON and 4xx are hinted control-plane errors; max_completion_tokens for api.openai.com vs max_tokens for compat servers (by host); an injectable FetchLike for deterministic, network-free tests.
• provider registry + resolveModel('provider:model' / bare provider / bare model name by prefix) + defaultModel() env auto-detect (OPENAI→GEMINI→GROQ→ZAI→ OLLAMA_BASE_URL→VLLM_BASE_URL); a missing key yields our hinted error, not a raw 401; Anthropic's native dialect returns a clear 'not yet ported' error.
• tests (injected fetch, no network): text + tool-call mapping, tool-calls-win, malformed-args error, only-the-intent-surface on the wire, 4xx hinted, token-field by host, and resolution (provider:model, bare-name sniff, ollama needs a model, missing key, anthropic not-yet, unknown provider, env auto-detect).

Verified: make js-check green — format, typecheck, 139 tests (138 pass, 1 skip).