refactor(worker): move job handler data access into business services

[realcodesiman]

Summary

• Removes every direct db.* call from apps/worker/src/{default,schedule,trigger,sequence-scheduler,webhook} (export/import handlers, tag sync, broadcast crons, quota reconcile, trigger engine, sequence scheduler, webhook engine) per .agents/rules/data-access.md. Bodies are moved verbatim; chunk sizes, orderings, onConflict targets and raw SQL are unchanged.
• One of seven parallel scope PRs following feat(contacts): cover full public API for MCP, move logic into business services #1093 (contacts); based on main, append-only barrel edits. The new contact repository also exists on feat(contacts): cover full public API for MCP, move logic into business services #1093; keep one copy when merging.

Changes

• New tag-channel and sequence-dispatch repositories, mac-partitions.ts DDL helpers, contactRepository.listForExportPage, fileRepository.updateForWorkspace, contactInboxRepository.{listByInboxPage,listContactIdsByIds}, integrationMessengerRepository.{findById,findByInboxId,listByWorkspace}.
• broadcastService: listDueScheduled, listSendingAwaitingHandoff, findScheduledForPrepare, resolveTemplateIntegrationMessengerId, insertRecipients, promoteAfterPrepare (keeps the resumeCount = promotionEpoch CAS guard and returns the boolean the handler branches on), listSendableById, listPendingRecipients, markContactFailed.
• userQuotaService.{reconcileUserSelfUsage,persistMacUsed,applyMonthlyBotMessagesReset} and workspaceUsageService.{loadReconcileCounts,upsertReconciled}: direct-assign semantics (never GREATEST), onConflictDoUpdate kept, self-path hset has no mac field, DB zeroing still precedes the Redis field write (AGENTS.md invariants 11 and 12).
• triggerService, webhookService, tagService, conversationService, flowService, inboxService, zaloIntegrationService: trigger-engine and sync-tag reads/writes. tagService.{attachExistingToContactForTrigger,detachFromContactForTrigger} deliberately do not emit or bulk-enqueue; the worker keeps its own per-pair enqueue loops. hardDeleteSoftDeleted keeps the isNotNull(deletedAt) guard.
• sequenceDispatchRepository.claim keeps the status='pending' predicate + affected-row check; deleteTerminalBatch keeps the CTE with the partition-pruning sd.workspaceId = rows.workspaceId join.
• contactService.insertImportedContactBatch (body in contact/insert-imported-batch.ts, exposed as a normal class method) holds the import transaction verbatim.
• 22 worker test files re-pointed to the service boundary; SQL-shape assertions relocated into business/database unit tests rather than dropped; new blocked-owner-guard and job-id regression tests.

Notes for reviewers

• isBlockedWorkspace is still called at the same points in prepare-broadcast and process-broadcast-contacts, and was not added to the fan-out crons.
• packages/business/src/tag/service.ts, flow/service.ts and the two barrels will conflict trivially with the sibling worker-integration PR (refactor(worker): move integration, chat and ai-agent handler data access into business #1099) and feat(contacts): cover full public API for MCP, move logic into business services #1093; all are append-or-keep-both.

Test plan

• pnpm --filter @chatbotx.io/database check-types && test
• pnpm --filter @chatbotx.io/business check-types && test (only the pre-existing ads-conversion-rule.service.test.ts failures remain, also red on main)
• pnpm --filter worker check-types && test (only the pre-existing flow-import-handler.test.ts failures remain, also red on main)
• pnpm --filter worker build
• pnpm lint
• Staging smoke: export + import a contact CSV; sync a tag to Messenger/Zalo; schedule and send a broadcast; run the quota reconcile cron once