Every single request llm() makes (one item, or one batch, depending on
the caller) is bounded by this wall-clock limit. Without it, a stuck local
model (most often Ollama under load, or a batch too large for the model to
finish) blocks silently – no log output, no progress, nothing – for as
long as the underlying HTTP client is willing to wait, which can be hours.
A timeout turns that invisible stall into a fast, visible, per-row failure
(recorded the same way a provider error already is – see llm()'s own
error handling), so a large batch run fails fast and retries instead of
hanging.