`src/engine/bare/llm_turn.rs` — one model call

Request construction and the two turn paths.

Key items

build_turn_request(transient) — extras first, then machine.full_history(); system prompt; tool schemas (None when the registry is empty).
turn_request_options() — your request_options + the fallback routed_model() override (which wins over your model field when routing is active).
do_turn — the cancel guard, then do_create_message (biased select vs cancel + turn timeout) or do_stream (through the StreamHandler; AttemptReset wipes the accumulator; Fallback adopted wholesale).
routed_model() / routed_or_client_model() — which model serves this turn.
record_turn_success / record_turn_failure — the fallback-breaker glue (rate-limit escalations count as RateLimit; trips fire on_fallback; chain advance on failed fallbacks).

Behavior notes

The user’s input reaches the model only via the scratchpad — turn_input is for memory retrieval and observers, never the request.
Overhead (system + schemas) is measured once with the configured counter; response_format turns suppress tools so their estimate over-reserves — conservative by design.
Detection feeds only on terminal replies (no tool calls).