feat/ques: multi-turn tool calling

IRL tools are usually designed to be chained. Would Needle, or its architecture, work for chains where there's only a minimal amount of stuff between tool calls (match -> reformat -> call -> repeat). I am really excited for these models, and hope that they can serve as a "first line" where larger models get less involved, and energy usage is massively reduced for simple AI use-cases that don't need any Intelligence, just Comprehension (I would say AC, but that was taken)

Start by reviewing Needle's architecture and any existing tool-calling or inference entry points; the issue does not name specific files or tests. Determine whether a match → reformat → call → repeat chain is supported and what interface it would require. Done is not defined because the issue currently asks an architectural question rather than specifying an implementation or acceptance criteria.

Written by the indexing model from the issue text.