Faithful port of DSPy 3.2.1 TwoStepAdapter (dspy/adapters/two_step_adapter.py).

Two stages:

1. The MAIN LM receives a simple natural-language prompt: a persona/task system message built from the signature's field descriptions (no [[ ## field ## ]] markers), demos rendered as plain name: value user/assistant turns, and a plain name: value user message.
2. Parsing runs a SECOND extraction LM through the ChatAdapter path over a synthesized text -> {original output fields} signature, extracting the structured fields from the main LM's free-form completion. A failed chat extraction falls back to a JSONAdapter-formatted retry, exactly like DSPy's ChatAdapter.__call__ fallback.

Extraction LM threading (mapping to DSPy's constructor argument)

DSPy carries the extraction LM on the adapter instance: dspy.TwoStepAdapter(extraction_model=lm). Imp adapters are stateless modules, so the instance slot maps onto Imp's settings surface — the same place dspy.configure(adapter=...) state already lives:

Imp.configure(adapter: Imp.Adapter.TwoStep, two_step_extraction_lm: small_lm)
# or scoped:
Imp.Settings.context([two_step_extraction_lm: small_lm], fn -> ... end)

A direct parse/3 call may instead pass extraction_lm: in opts. A missing extraction LM is a loud error: parse returns {:error, {:two_step_extraction_lm_not_configured, ...}} rather than silently degrading to single-step parsing.

The former plan-prepend extension that wore this name is now Imp.Adapter.PlanFirst.