Prevent stop marker from mutating OpenAI-compatible prompts

[jxnding]

Summary

• avoid mutating caller-owned chat turns while formatting GPT requests
• apply stop-marker prompt shaping only to the OpenAI Responses API path
• add regression coverage for custom OpenAI-compatible and Responses requests

Root cause

sendRequest previously ran strictFormat(turns) and appended stop_seq before choosing an API path. Both operations mutate the message objects. When a custom URL selected chat.completions, that branch reused the same turns, so the stop marker was sent inside the prompt as well as in the stop request option.

Some OpenAI-compatible servers can interpret a stop marker already present in the prompt as an immediate stop and return an empty completion. The mutation also leaked back into Mindcraft's conversation history and could accumulate across retries.

Fix

Clone messages before passing them to strictFormat, and append the prompt stop marker only in the Responses API branch. Custom endpoints continue to receive the existing stop option, while their message content remains clean.

Validation

• npm test on Node 18.20.8, 20.20.2, and 22.23.2
• two regression tests covering custom endpoint payloads, caller immutability, Responses prompt shaping, and output truncation
• manually reproduced the empty response with a llama.cpp OpenAI-compatible endpoint before the fix and confirmed a non-empty response afterward