Responses→chat.completions shims (observed: gpt-5.x behind
OpenAI-compatible relays) split each reasoning-summary part across two
fields: the bold headline arrives as a single delta.reasoning_content
chunk (**…**) while the part's BODY streams as plain delta.content. The
planning prose then rendered as the assistant's reply, concatenated
across steps without separators.

Provider stream loop now runs a small state machine: after a
headline-only reasoning chunk, content deltas buffer; a tool-call delta
or finish=tool_calls proves the buffer was planning prose → flushed into
the reasoning block; finish=stop or outgrowing the 4KB summary cap
proves it was the real answer → flushed back as ordinary content.
DeepSeek-R1-style multi-chunk reasoning prose never matches the headline
gate, so real answers keep streaming live. Summary parts join with blank
lines; the chat reducer also inserts a paragraph break when a new
headline lands flush against the previous step's reasoning.

Verified against the live relay wire shape (headline via
reasoning_content → body via content → tool_calls) and covered by six
new offline stream tests.

…out finish_reason