[UsatovPavel]

Jira

HACK-186

Production incident

The ML v4 deploy on main started the equipment inference image before the Backend migration. Equipment warmup completed, but each tick then failed with column "journal_is_alarm" does not exist; the container remained starting and stopped producing predictions. The deploy run has the protected 72-hour snapshot already saved.

Fix

• Equipment queries use a typed NULL for the incident-only raw alarm column. Incident queries still read the real column after Backend migration.
• The mandatory pre-switch backup can continue when the old container is running and its recent logs prove this exact missing-column failure. It still copies and verifies the cursor, model, reference views and 72-hour inputs before switching images; other unhealthy states block deployment.

Verification

• python -m pytest -q: 70 passed.
• bash -n deploy/backup-equipment-window.sh.
• Added regression tests for the pre-migration equipment schema and guarded recovery backup.

Merge and deploy this hotfix before Backend #66. Confirm inference becomes healthy and writes fresh equipment predictions, then proceed with incident rollout.