fix(llmcore): handle Responses API incomplete/failed terminal events + reasoning_text

[amm10090]

…+ reasoning_text

• response.incomplete/response.failed now record usage, emit truncation marker and stop cleanly instead of being misread as ConnectionError (retry storm)
• collect response.reasoning_text.delta/.done into thinking content blocks
• non-stream Responses JSON: parse reasoning items and failed/incomplete status

[lsdefine]

Reviewed and verified locally by replaying simulated SSE streams against this branch:

• response.incomplete/response.failed with no prior output previously fell through the loop end -> empty blocks + no yields -> ConnectionError("empty response") retry storm (llmcore.py L415-416 on main). This PR fixes the real root cause; usage is recorded and a terminal marker/!!!Error: is surfaced, consistent with existing conventions.
• reasoning_text collection mirrors the chat_completions path (thinking blocks), and non-stream reasoning item parsing checked out.
• Style matches the codebase; no unrelated changes.

One minor note (not blocking): when partial text was already streamed before response.incomplete, no truncation marker is appended — conservative and acceptable.

Merging. Thanks for the clean, well-scoped fix!