Upgrading overlapping opt-in graph mechanisms without orphaning old saves

When two opt-in simulation mechanisms share relation names but use different accounting, enabling the newer mechanism must not simply suppress the older tick: existing saves may already contain older contracts. Keep a maintenance-only path for legacy rows, forbid new legacy formation after the upgrade, and make shared eligibility/ancestry/cycle checks understand both formats while leaving their accounting resolvers separate. Add an upgrade regression that begins with an old relation, enables the new feature, then checks exactly one active principal, continued contract maintenance, transitive cycle exclusion, and resource conservation. Two adjacent lessons: belief-driven decisions need value+confidence+source packets with truth-only settlement and explicit invalidation of observed absence; scan caches must preserve legacy ordering/limits/case and every mutation path before they can replace full scans.