Backup failing on Mac (formerly WIN Katalog)

Problem: Lightroom Classic catalog backup fails every time, consistently at the final step — "Katalogsicherung wird komprimiert" (compressing the catalog backup). Error shown: "Lightroom konnte den Katalog '2026_07_LightroomFull-v14' nicht sichern" with generic advice to check folder permissions/disk space.

System:

• Lightroom Classic 15.4.1 [202606201310-b9f148a4], Creative Cloud license
• macOS 26.5.2 (25F84), Apple M2 Max, 32GB RAM, arm64
• Catalog:  (~4.2GB), on an internal HFS+ (Journaled) volume connected via PCIe
• Backup target tested:  (on boot APFS volume) — also failed with 2 other previously tried destination folders

Already ruled out:

• Destination folder permissions/ownership — correct, and writable (verified via Terminal)
• Free disk space — 213GB free on boot volume, 4.2TB free on catalog volume
• macOS Full Disk Access — granted to Lightroom Classic
• Filesystem corruption — Disk Utility First Aid on the catalog volume: clean, exit code 0
• Catalog SQLite structural integrity — PRAGMA integrity_check on the .lrcat file returns ok

Key finding: Lightroom's own log at get_table_failure.log recorded repeated "Blob DB service is not initialized" errors on 2026-07-15, six days before backups started failing (first failure logged 2026-07-21, still failing as of today, 2026-07-25 — confirmed via big_data_backup_file.log, which logs a "Backup failed" entry on every attempt with no further detail). The Blob DB is the AI-edit data store (.lrcat-data folder, used for Denoise/masking/AI features). When testing whether this store was missing/rebuildable, Lightroom explicitly warned that proceeding without it would invalidate all existing AI-based edits — confirming it holds live edit data, not disposable cache, so it wasn't safe to continue investigating by removing it.

Question for Adobe: What causes "Blob DB service is not initialized," and is there a supported/safe repair path for this store that would let catalog backups (specifically the compression step) complete without risking loss of existing AI edit data?