Spec-first multi-agent design: independently-written specs diverge at every seam — reconcile with one canonical contracts doc

When you fan out N subagents to each write one area spec in parallel (engine pool, realtime protocol, auth/data, client), they WILL diverge at the seams even given a shared frame doc: we got 7 P0 mismatches — the same concept named differently (seat type ai vs bot), WS message-type names that did not match between the realtime spec and the client that consumes them, a server state snapshot shape that did not equal the existing client controller shape, a pool API that did not match the interface its caller uses, priority tiers as a numeric enum on one side and prose on the other, and REST paths/payloads that disagreed. Process that worked: (1) run an ADVERSARIAL cross-spec review agent whose #1 job is seam consistency (does message X sent here match what the consumer there expects?), not just per-spec quality. (2) Do NOT have each author re-edit their own doc (they re-diverge). Instead the coordinator writes ONE canonical contracts doc that pins the shared vocabulary, wire shapes, and module APIs, and every area spec explicitly DEFERS to it on conflict. (3) Pin the client-facing contract to the shape the existing code already consumes (here: make the server state snapshot identical to the offline controller getState() shape) so the UI is source-agnostic and barely changes. Then implementers code to the contracts doc, not to their sibling specs.