Rebuild Babylon Showroom analysis

This is an opt-in maintenance operation, not a deploy/startup migration. Do not run it without separate approval for the selected environment and content IDs. Production is not authorized by this change. Never use init-db --drop, table drops, or truncation for this rebuild.

What changes

For the enumerated Babylon IDs only, reset generated Showroom analysis fields, generated entity-card fields, summary/module embeddings, scan state, and overlap candidates involving those IDs. Analysis rows remain in place to preserve curated duration, notes, enrichment-review flags, and review reasons. Source provenance, curator URL/path overrides, catalog identity and metadata, retirement state and workflows, Jira/GitHub links, usage/performance/nonprod/field-source records, advisor/job/token history, enrichment, infrastructure/workloads, vocabulary, credentials and roles are not reset. OSSPA analysis/embeddings are untouched. Ordinary non-Showroom sandbox cards come from the infrastructure-summary workflow; they are excluded from this reset and its affected-item counts, even when selected by --all-babylon.

Maintenance prerequisites

1. Obtain approval for the API-only dev deployment; deploy the complete isolation change from an approved pushed feature ref or merge. Never deploy partial implementation slices or uncommitted binary builds. Verify additive schema columns and a small fixture set before any bulk reset.
2. Record preservation snapshots/backups securely outside the repository: selected identities, human Showroom fields, retirement state/workflow links, source overrides, usage records, advisor/job/token history, and OSSPA counts and artifacts. Keep a record of the selected IDs and dry-run counts.
3. Stop scheduled and manual scan producers: disable maintenance scheduling and block refresh/scan/rescan endpoints and maintenance submissions for the window. Stop all scan-worker replicas, including any local development workers. Stop and wait for all CLI scans, catalog refreshes and sandbox-summary processes. Coordinate with other operators; a replica count of zero does not stop a CLI.
4. Drain/discard previously queued and in-progress scan work from Redis/arq and reconcile the corresponding DB job states using the existing operational tooling. Preserve job records; do not delete history. Confirm no worker or thread can still persist a pre-reset result, including timed-out CLI processes.
5. With producers still blocked, refresh catalog source metadata using the new resolver (rcars refresh) and let that command finish. Inspect invalid and ambiguous sources; they are failures, not ordinary no-Showroom sandboxes. Reconfirm quiescence and drained work after refresh.

The reset rejects outstanding queued/running scan-related DB jobs. That check cannot detect every CLI process or prevent a new producer race. --confirm-scans-paused is an operator attestation to the prerequisites above, not an automatic proof that the system is race-safe while producers are running.

Dry run, review, then execute

From src/api, with RCARS_DATABASE_URL securely supplied in the environment (do not write endpoints or credentials into this document):

python scripts/rebuild_babylon_analysis.py \
  --content-id babylon:example.content.prod \
  --content-id babylon:example.virtual.dev

The default is read-only. Review the exact IDs and counts, preservation snapshots, and approval before adding both execution flags:

python scripts/rebuild_babylon_analysis.py \
  --content-id babylon:example.content.prod \
  --content-id babylon:example.virtual.dev \
  --execute --confirm-scans-paused

For an approved full Babylon rebuild, use --all-babylon instead of repeated --content-id; this includes retired IDs so contaminated artifacts do not remain behind. It still leaves identities retired and does not scan them.

Rebuild and verify

Keep ordinary producers blocked until the controlled rebuild is finished. Use the existing scanner for the reset targets, not a second scanning implementation:

rcars scan --content-id babylon:example.content.prod \
  --content-id babylon:example.virtual.dev

For the approved full rebuild, rcars scan processes every eligible active stage that needs work, including virtual items. Do not add --force routinely: unchanged complete items skip jobs/LLM/embeddings; content-identical commits only advance the verified revision; embedding-only changes repair affected artifacts. --force is available when separately requesting analysis from scratch.

Verify LB2081b/c fail with insufficient/template-only content rather than Ansible analysis, while LB2081a retains correct PQC analysis. Check differing stage refs, a multi-hop virtual chain whose content component has no CatalogItem, independent same-URL/hash items, fresh target-specific embeddings, and source/signature/donor provenance. Compare all preservation snapshots and ensure failed scans cannot search the old summary. Retired items must remain retired.

After all scans finish, recompute overlap candidates/assessments using the existing overlap workflow. Confirm no old scan jobs remain, then restore scan workers, manual producers and maintenance scheduling. A second routine run should submit no unchanged analysis jobs. Rollback must use reviewed backups; do not restore contaminated derived artifacts automatically or discard historical records.