Byte length detected by the file adapter.
Ordered CSV column names.
Array position is the zero-based column index. Duplicate names are allowed because duplication itself can be a useful intake or data-quality signal. When the source header is blank or absent, the adapter should provide a deterministic placeholder instead of asking the model to invent one.
Compact handoff memo for later agents.
Explain what the content item appears to contribute, which downstream phase should care, and what must not be inferred yet. For example, an ERD can be structural, a meeting transcript can be requirement-oriented, and one item can be both when annotations contain business rules beside structure.
Number of data records in the CSV, excluding the header row when present.
This is adapter-owned metadata. URL-backed files should be streamed by logic to count records exactly; downstream workers use bounded chunks only for model-facing evidence, not for metadata counting.
Metadata branch discriminator.
Downstream semantic roles for this content item.
Empty means the content item should not feed a phase directly. Supporting evidence, metadata-only value, policy limitations, intake gaps, or unsupported content should be explained in memo instead of adding a separate enum here.
CSV-specific metadata attached after deterministic CSV file intake.
This branch carries the structural facts that later ontology workers should not ask the model to rediscover by loading the whole raw file. It is produced by CSV intake, not by the facade semantic metadata function call.