Decouple state machine and context from LLM modules

[20001LastOrder]

The core of the modernization. Split Sherpa into (a) a state machine + context management core that owns workflow, transitions, guardrails and what-to-keep/what-to-throw, and (b) the LLM modules that currently do generation, tool calling and parsing. Once separated, the LLM half can be supplied by a host agent — Claude Code, Codex, or another — and Sherpa becomes the explicit-workflow layer they plug into.

Everything else in this project is downstream of this split.

• Draw the seam: what the state machine and context layer actually need from an LLM, expressed as a narrow host interface (propose a step, judge a condition, produce structured output)
• Identify every place the core reaches into model wrappers / policies / providers today
• Extract the core so an external caller can step it one transition at a time
• Keep the current in-process LLM modules as the reference host implementation — no behaviour change for existing users
• Standalone mode has to keep working, or the decoupling costs Sherpa its current users