Attribute promotion and expansion
baselode.datamodel.attributes.promote_attributes adds explicitly mapped text fields to unique parent records without changing existing columns or row order. Long-form attribute records must supply parent_id, attribute_id, name and value; an optional modified_date supports ranking. Additional scalar record metadata is retained.
Mapping entries declare name, target and integer priority. Names normalize camel boundaries and punctuation to lowercase snake_case. Lower priority wins, then newest modification date and lowest attribute ID. Null or blank values cannot win a promotion, but remain in source evidence. Unknown targets and normalized mapping collisions are errors.
attributes_extra is UTF-8 JSON text, suitable for Parquet and JSONB ingestion. Unmapped normalized keys contain an original scalar value or distinct-value array. Null values and normalization collisions remain represented. _attribute_records retains every source attribute record, including original names, values and metadata. _conflicts contains ranked candidates for mapped fields with differing names or usable values, with the selected candidate identified. Alias differences count as clashes even when values agree. attribute_clash_count counts clashing output fields.
Null/duplicate parent keys, null/duplicate attribute IDs and orphan attributes are rejected. No interval correction, scientific code translation or numerical imputation is performed. Mapping policies belong to the caller; the primitive does not infer synonyms from prevalence.
Mapped attribute values must be text or null; unmapped values may use any supported JSON scalar encoding. Numeric source formatting must be preserved explicitly before constructing the input DataFrame. Priorities must be integers; fractional, boolean and string priorities are rejected. Conflict candidates contain a nested source_record, so source metadata cannot collide with selection bookkeeping. json_scalar supports JSON scalars, ISO dates/timestamps and tagged Decimal, bytes and infinite-float metadata; unsupported types are rejected.
Expansion output names are reserved. Numeric attribute IDs sort numerically; text IDs use natural numeric-segment ordering, with original text as a final tie-break. Mixed numeric/text or other ID types are rejected. Modification dates for mapped candidates must be explicit ISO calendar date/time text (basic or extended) or date/timestamp scalars, or null; numeric epoch values and unparseable dates are rejected with the offending record ID. Unmapped modification metadata is retained without being interpreted for ranking.
Date-only or timestamp strings must include the full calendar date. Relative tokens and partial dates are rejected so candidate ordering never depends on the execution clock.
Mapped dates must fit pandas nanosecond ranking precision (approximately years 1677–2262). Dates outside that range are rejected with their source record ID, consistently across pandas versions. Original unmapped metadata remains evidence and is not ranked.