Capabilities : What a migration connector declares it can do. The engine consults this before it starts, so an unsupported operation is refused up front instead of failing part-way through a run. Every field defaults to false for a connector that has not declared it – an absent capability always reads as "no", never as an optimistic "yes". There is deliberately no per-field "unknown" state: unlike a write-permission probe (which can genuinely fail at runtime, and where the engine branches on the difference), a capability is a static statement a connector makes about itself, and "not declared" and "cannot" produce identical behaviour. Per connection, not per connector type – two instances of the same connector may answer differently (a DuckDB store over Parquet has a schema, one over a headerless CSV does not).