The LLM client's literals read at the point of use, and its test servers answer politely

[WaylandYang]

Follow-up to #538, no behaviour change in the client.

• Every literal that LLM errors preserve their response bodies #538 had pulled into a named constant is back at its point of use: "LLM", "Embedding" and "unknown error" in the client, and the twenty test fixtures ("hi", "user", "502 Bad Gateway", "界", the number one). A name earns its place when the value means something beyond itself or has to stay in step across sites; none of these did, and in a test the literal is the documentation. MAX_ERROR_DETAIL_CHARS keeps its name and gains the reason for 500 and for counting characters.
• response_failure and err_detail get doc comments in this repository's voice: what changed, the LM Studio 400 that motivated it, and why a body that cannot be read is still Unreachable.
• The six tests and three helpers get one-line comments saying what each pins, and the Unreachable doc sentence reads in one breath.
• The raw-socket test servers now read the request (headers, then the body by Content-Length) before writing their answer and shut down cleanly. Written as they were, they closed the socket with the request unread; Windows answers that with RST and the client sees "connection aborted" (os error 10053) instead of the response, so two of the six tests failed on a Windows checkout while passing on the Linux CI. Same behaviour under test, now the same on both.

cargo test -p utopia-llm passes, three runs in a row on Windows.

🤖 Generated with Claude Code