feat(ai-provider): messages.stream — the adapter half that was missing #5627

Summary

PR B of the multi-provider build. AiClientLike implemented only messages.create, so AI_PROVIDER=openai crashed all four streaming features (personal-chat, repo-chat, spec-ai, streaming-review) with messages.stream is not a function.

• One SSE pump drives the connection regardless of consumption style, serving both styles the features use: for await over Anthropic-shaped events (content_block_delta / message_delta with usage / message_stop) and .on("text") + .finalMessage().
• stream_options.include_usage asks for token counts on the tail chunk so cost tracking sees real numbers; servers that ignore it degrade to 0, never NaN.
• Failure is recorded before the iterator wakes — a truncated stream ends loudly, not cleanly. Torn SSE frames skip; frames split across reads reassemble.

With this merged, one env var (AI_PROVIDER=openai + AI_BASE_URL) points the entire platform — streaming features included — at OpenAI, Azure, OpenRouter, Groq, Together, Ollama, vLLM, or llama.cpp.

Test plan

• 6 new tests: both consumption styles, finalMessage-only drain, non-2xx rejection in both styles, torn frame, split frame — pass
• Existing ai-provider suite green (18 total), tsc --noEmit clean, bun build clean