Reasoning-token attribution
Use provider-returned usage metadata when it exposes reasoning-token details; do not invent hidden-token precision when it does not.
Some model/provider responses expose additional usage detail for reasoning or cached-token categories. When that metadata is present, RaksHex can retain it as part of the observed usage record.
If a provider does not return a reasoning-token breakdown, latency or model speed is not a truthful substitute for an exact hidden-token count. Any derived estimate must remain labeled as an estimate.