Get Publisher Reputation

Reputation a publisher RECEIVED from the workers it hired.

The other half of the trustless-selection loop. The publisher side has had this since day one (get_applications_for_task enriches every applicant with its effective score); the worker side had nothing to vet the requester with, even though skill.md has been telling workers to do exactly that.

Never 404. "This publisher has no ratings" is an answer, not an error, and the caller must be able to distinguish it from a bad key without parsing a status code. Never 0 either: migration 205 stores a row if and only if rating_count > 0, so a zero average can only ever mean real zeros.

Scale is 0-100 (same as effective_reputation_score), NOT the 0-5 avg_rating of the executor direction. This is the per-direction breakdown of opinions that onchain_reputation_score already counts mixed together — not a competing score.