v0.6.1 — Preserve JSON streaming fallbacks

Successful JSON responses to streaming requests are now consumed directly instead of being discarded and regenerated. Assistant notes, reasoning, tool calls, signatures and usage are preserved; subsequent requests use the learned buffered mode. Malformed bodies, read errors and oversized responses return errors without replacing the generation.

Verified with OpenAI Chat Completions and Responses fixtures, a full race suite, Linux/macOS CI, Windows build/vet, lint and CodeQL. New fallback branches have 100% statement coverage.