dirigent-block-parquet

The parquet block family: one codec, convert.arrow, trading parquet with the three text spellings convert.std already trades between.

The record model is convert.std's: a sequence of flat records. What parquet adds is types, and the codec is honest about them in both directions:

• Writing infers one type per column -- boolean, int64, float64, or string -- from the records. Integers and floats unify to a float column, because JSON calls both a number; any other mix is refused naming the column. A column every row leaves null is written as a nullable string column. A csv source carries no types, so csv to parquet writes string columns and nothing else.
• Reading gives every value its JSON spelling: timestamps, dates and times come back as ISO strings, decimals as strings, and a float that is NaN or infinite as null. A nested column, or one holding raw bytes, is refused naming it -- flattening is a reshape, and a reshape belongs to a jq step that knows what the flattening should mean.

A conversion is a storage-object operation, like storage.copy: it reads the object at source and writes the one at target, and neither side is ever a value the step carries. A run that means to look at the records reads them with storage.read.

The family ships separately because convert.std is deliberately on the standard library and nothing else; this codec stands on pyarrow. Nothing else in the workspace may depend on it.