BOLT-1268 P1 — address Bugbot HIGH: publishedAt write failure window on PR #12181

New Cursor Bugbot HIGH finding on PR #12181 thread PRRT_kwDOJpxUwM6JNiEH: if executeBatch succeeds for some batches but the post-loop workspaceProgramEventRepository.publishWorkspaceProgramEvent (DB write to set publishedAt) throws, devices have received commands but publishedAt stays null. User retry passes idempotency guard and redispatches everything — same class of bug PR fixes for Edit→Schedule. Fundamental constraint: executeBatch is a network side-effect to supergraph; cannot be atomic with a Prisma write. Real fix is per-batch dispatchedAt tracking on Batch model (file as separate Linear). This PR scope: add a retry wrapper with bounded attempts (e.g. 3 attempts with exponential backoff) around the publishedAt DB write. If after retries it still fails, log Datadog error explicitly identifying the half-state with the workspaceProgramEventId so on-call can manually flip publishedAt. Add a code comment documenting the residual risk window. Add a test for the retry behavior. PR #12181 / branch winston/bolt-1268-p1-idempotent-publish.