Adversarial review of independently-written backend specs: check the seams, not the specs

When 4 backend area-specs (WS realtime, engine-pool, auth/DB, client) are written independently against a shared architecture frame, the bugs are almost never inside a spec — they are at the seams. Highest-yield checks, in order: (1) message/field NAMING drift across specs sharing a wire protocol (found gid vs gameId, auth/hello vs welcome/auth-ok, error code illegal vs illegal_move) — grep each message type across all specs. (2) The claimed contract vs the ACTUAL code it must match: read the real snapshot()/state-shape and diff field-by-field (found clock wire shape invented a ms field that PlayerClock.snapshot never emits; seat type ai in code vs bot in specs). (3) Method-name parity between a new pool/adapter and the existing interface it claims to implement (pool.analyze vs GoAI.analyzeGame; scoreFinal returning deadStones that no producer emits). (4) Reuse claims: open the file being reused — a spec said it reuses katago-bridge as a driven library, but the file is a standalone WS+http server with an unexported class; another claimed verbatim reuse of query builders that are private methods on a WebSocket-coupled class. (5) A single coarse server-side clock tick + immediate turn-handover mis-attributes since-last-tick think time to the opponent — must tick the mover before handing over. (6) A referenced companion spec that simply doesnt exist in the review set (matchmaking) leaves double-accept races unowned. File one issue per seam with exact file:line on both sides.