StaticBuildImpute (with a missing indicator) every column, one-hot every categorical, then the pipeline's explicit ops.
StaticColumnThe names of the matrix columns the assembly writes for these feature columns, in order: one per column, one per
date part of a date column (<name>__<part>), none for an identifier. The assembler names its columns with
ColumnsFor, so a caller that needs the names before assembling (training checks a protocol's slice fields
against them) gets exactly the assembled names.
StaticColumnsThe matrix columns one feature column is written as (its date parts for a date column).
StaticContributorsA windowed aggregate's source version: the sha-256 of the contributing rows, ID:updatedAt per row (or the ID
alone when the row has no update time), sorted, one per line. The one rule for the version, so a reader that
rebuilds the contributor set can check a stored version against it.
StaticToAdapter for the training step and the sidecar request: matrix + schema + preprocessing. With labels the label is
appended as labelColumn (training); with null no label column is added (scoring).
ReadonlyCategoryLevels: ReadonlyMap<string, readonly string[]>The kept levels of every categorical column, by column name: persist them with the model and pass them back at scoring.
ReadonlyColumnProvenance: readonly ColumnProvenance[]ReadonlyColumns: readonly AssembledColumn[]ReadonlyCoverage: readonly ChangeHistoryCoverage[]ReadonlyExcludedRowIDs: ReadonlySet<string>ReadonlyPreprocessing: readonly PreprocessingOp[]ReadonlyRows: readonly AssembledRow[]ReadonlyStatus: "assembled"ReadonlyUncoveredRowsExcluded: readonly UncoveredRowsCount[]
Assembles the dated matrix, or the first refusal met (caught with
instanceof, never by message).