# Public API
`DEPTH_DECAY` is `0.7`. `score(value, depth)` returns `value * 0.7^depth` in constant work.
`choose(candidates, temperature, random)` returns an index within a nonempty candidate slice. Each
candidate is `(frontier_index, effective_score, attention_cost)`; selection reads only the score.
Callers provide finite positive scores and a finite nonnegative temperature. Zero temperature chooses
uniformly among maximum-score ties. Positive temperature samples weights proportional to
`score^(1 / temperature)`. The supplied random operation is invoked once and must return a finite
value in `[0, 1)`. Selection is linear in candidates and uses temporary memory linear in candidate
count; callback completion has no finite wall-clock bound.
`os_random_unit()` returns a uniformly constructed `f64` in `[0, 1)` from one operating-system
entropy fill. Entropy failures retain their complete child message. OS completion has no finite
wall-clock bound.