fix(llm): surface server error bodies instead of silent NOTFOUND

[vxfemboy]

Problem

When an OpenAI-compatible server returns an error body at HTTP 200 — e.g. mlx_lm.server given an [ask].model it doesn't have loaded returns {"error": "404 … Repository Not Found for kuro"} while trying to fetch it from HuggingFace — kibble silently swallowed it:

• chat_turn read v["choices"][0]["message"] (null) → content = None.
• chat_turn_stream's non-SSE error body matched no data: line → empty.

Both looked like an empty completion, which ask classifies as NOTFOUND → the user sees "Searched the corpus — nothing covers this" with zero indication the LLM call actually failed. (Cost me a long debugging session to trace a mis-set model name.)

Fix

• Add server_error() — pulls the message from an {"error": …} body (string or {message} object).
• chat_turn: return a clean Err("LLM server error: …") when the body is an error.
• chat_turn_stream: after the loop, if nothing streamed and the leftover buffer parses as an error body, Err the same way.

Now a bad model / auth / 404 surfaces the real message instead of a phantom "not found in corpus."

Tests

• chat_turn_surfaces_server_error + chat_turn_stream_surfaces_server_error (buffered + streamed error bodies → Err), verified RED before the fix.
• Live-verified: kibble ask with a bad [ask].model now prints the HF 404 instead of "nothing covers this."
• Full suite 381 + integration, cargo clippy --all-targets -- -D warnings clean.

🤖 Generated with Claude Code