[ewiger]

What

The analytic windows of HQL-0050, and a total order for keyed selection. over partitions a sequence by a key, orders each partition, and gives one value per row, at the row's input position, from the row and its Window:

ledger
| over(by = t => t.account, order = t => t.seq,
       into = (t, w) => w | running | map(x => x.amount) | sum)

• over(by, order, into) in seq: order is a key or a comparison; rows equal under it keep their input order. No row is merged or dropped.
• Window<T> in core: row_number, rank, dense_rank, lag, lead, first, last, size, and the frames running, rows(before, after) and whole. Called as a step, w | row_number, or as a receiver, w.row_number().
• then in collections: compare(a.at, b.at) | then(compare(a.seq, b.seq)), the second comparison evaluated only on a tie. A comparison that ends in a unique key is a total order, so keyed selection does not depend on input order.
• Refused at check time: an into whose result holds a Window, and over of a Stream or a Set.
• sort is specified as stable, which it already was.

Breaking

Window is a standard type name, so a program's own type Window is refused, as Group and Pair already are. The structural examples and the type-contracts page call their contract Interval now.

Not in this PR

HQL-0050 is implemented in part. Carrying input position across splits, and keyed selection as a partial reduction, wait for a backend that splits; localhost does not.

Docs

A new book page, Part II's "Windows" under OLAP, a tie-rule paragraph on the Deduplication page, generated reference pages, the changelog, and HQL-0050's spec as built.

Testing

examples/windows/ holds nine cases and two refusals, and examples/pipelines/queries/ranked-deduplication.hql runs the shop's replay through over. The cases were written as proposed with hand-computed results before the implementation, then promoted. cargo test --locked (653), cargo clippy --locked --all-targets, cargo fmt --check and hql reference --check pass locally; mkdocs build --strict was not run locally (no mike plugin here).