Finalize the visible text of stored messages

Record what a human actually saw for messages this conversation already stores. A chat client can only compute an assistant message's final visible text after the stream closes and its own parsers, postprocessors, and plugins have run; this is the channel for sending that back, so the transcript reopens showing what the visitor read rather than the model-channel payload. It is projection-only: it can fill or replace `displayContent` and nothing else, never creating a message or touching model content, role, ordering, timestamps, count, ownership, or title. The batch applies all-or-nothing, and the write is guarded by a compare-and-swap on `conversationRevision`, so it can never overwrite a turn the chat path persisted concurrently. Replaying an identical projection is a no-op that performs no write and returns the same `conversationRevision`. A genuinely changed projection advances the revision and refreshes the list `preview`, but deliberately does NOT change `updatedAt` — a closing tab must not reorder the visitor's history. Authorization is exact: the live session must be bound to this conversation, on top of the visitor secret. Small and bounded on purpose, so a client can send it with `keepalive: true` on page exit.