Inline (synchronous) interpreter for the pure Normandy.Agents.Turn core.

Drives a turn to completion in the calling process: it feeds :start into Turn.step/2, performs each returned effect's side-effect, and feeds the resulting event back into step/2 until the turn reaches :stopped (returns {:ok, state}) or :failed (returns {:error, reason, state}).

This is the library / scripted-run shell. It does NOT (yet) replace BaseAgent.run/2; it exists to prove the FSM core runs a real turn against a real Dispatch chokepoint. Compaction at the :steering boundary is supported via the optional :compact dep (default no-op); streaming, guardrails, validation, persistence and approval shells come in later phases.

deps is a map of side-effecting functions:

• :call_llm — fn request -> {:ok, response} | {:error, reason} end (required)

• :dispatch — fn calls -> [%ToolResult{}] end (required)
• :append — fn role, content -> any end (optional, defaults to no-op)
• :emit — fn name, meta -> any end (optional, defaults to no-op)
• :convert — fn raw, response_model -> converted end (optional, defaults to identity:

            returns `raw` unchanged)

• :validate — fn value -> validated end (optional, defaults to identity: returns value

            unchanged)

• :guard — fn value -> any end (optional, defaults to no-op: side-effecting, return

            is ignored and the input `value` is fed forward via `{:output_guarded, value}`)

• :compact — fn info -> any end (optional, defaults to no-op). Invoked at the

            `:steering` boundary; the inline shell does not thread memory, so
            a real impl must compact external state by side effect.

By design, step/2 always places the single blocking/terminal effect (:call_llm, :dispatch_tools, :finalize, :fail) last in its effect list, so the interpreter performs the leading :append_message / :emit_event effects in order and then acts on the terminal one.