v1.1.7

Fixed

• Rebase the chat message-id counter when a transcript is restored. The counter restarts at zero each page load, so the first new turn could reuse a restored message's id, producing duplicate React keys and causing a single update to patch two messages at once.
• Apply the same transcript flush and switch guard to snapshot restore that normal corpus switches use, so restoring a snapshot owned by another workspace can no longer persist an empty transcript over that workspace's saved history.
• Restore a saved transcript that arrived while an answer was still streaming, instead of discarding it permanently and letting the next save replace the stored history with only the new turn.
• Stop retrying a document-AI request that was cancelled during a rate-limit backoff, which continued issuing billable requests after the user had moved to another document.
• Mark an empty-bodied streaming failure as an error so it is excluded from the history sent to the model rather than replayed as a genuine prior answer.
• Stop re-reporting watched-folder files that were merely deferred by the batch size cap; they are retried on the next scan, so each poll was adding another ignored-file entry and warning toast until the backlog drained.