Tokens (inputTokens, outputTokens) now come from ACP PromptResponse.usage
which is populated synchronously from run_conversation() — no WAL race.
Turns still come from DB before/after snapshot.
Previously both were read from hermes state.db after ACP prompt returned,
but WAL write lag caused incomplete token data (e.g. 235 vs actual 26,080).
Refs #91
