Dataset migration and compatibility

Migration is projection, not relabeling. Preserve the original store, produce a mapping and loss report, verify the result, and keep rollback available.

Existing surface	Dataset compatibility path	Known boundary	Status
`index.graphs` and user indices	keep using existing index commands; legacy layout migration can build named sidecars	indexes/caches are regenerable, not canonical datasets	active compatibility workflow
semantic-memory `memory-ingest`	retain its summary, entity, contradiction, and topology behavior; optionally register source/run evidence	raw dataset ingestion cannot reproduce semantic synthesis	active; no retirement scheduled
research provenance	project records through the canonical ledger when useful	arbitrary entity/activity attributes require a governed extension	projection shipped; workflow remains active
marketplace provenance	project graph entities, activities, agents and relations	non-core entity attributes require a governed extension	projection shipped; workflow remains active
mention edges	project file entities, locator evidence and inferred assertions	a mention is not verified causality	projection shipped; workflow remains active
SDLC traceability	project requirement links with evidence and run identity	verified without a run ID remains inferred	projection shipped; workflow remains active
Fortemi v2 export	project field provenance into ledger events	search is not provenance; unsupported relationships are semantic loss	projection shipped; static export remains active
ledger to dependency graph	generate legacy graph view	drops basis, evidence, run, field, privacy and retention	lossy compatibility projection

Migration procedure

1. Back up or snapshot canonical sources and current configuration. 2. Run the existing workflow’s read-only check or dry run. 3. Record source revision, schema, privacy and retention. 4. Generate the compatibility projection and inspect every loss item. 5. Build beside the existing derived store; never overwrite the only copy. 6. Verify identities, counts, relationships, evidence and representative queries. 7. Switch consumers only after review; retain the old path for the support window. 8. Roll back consumer selection if verification fails. Preserve failed output for diagnosis.

No surface in this table is newly deprecated by Dataset Intelligence Stage A. The machine-readable deprecation inventory is empty until a release owner supplies a date, support window, automated detector, rollback, and measurable removal criteria.