DDIR: adopt corgi's copy-audit API (frankmcsherry/WIP#18)

[frankmcsherry]

Bumps the corgi pin from #16 (578ef3f) to frankmcsherry/WIP#18 (de0f2ac), which also picks up #17, and adopts the API changes that came with it.

Net −12 lines: three of the four changes delete code at the call site, because each one removes a thing DDIR was doing to work around a corgi API that made the caller pay for something.

What changed upstream, and what it does here

corgi::hash returns the ids. It returned a Value that every caller outside corgi's op vocabulary unwrapped in the same expression — ours was literally corgi::hash(col).into_u64("ids").unwrap(). Four sites lose the unwrap.

arrange::hash_rows is gone. It was a second implementation of the structural content hash, with different constants, and the two disagreed about whether a u8 5 and a u64 5 are the same value — which is exactly the question the boundary id exists to answer. Both were public and both documented as the boundary id. This repo had already worked that out: reduce.rs routes real ids through corgi::hash and calls the other one "branch-local". The only surviving uses were two lines of a round-trip test comparing the function to itself, so nothing observable moves.

Bounds::Offsets holds an Arc. A Value clone no longer copies a list's partition — which for a List<U64> was as many bytes as the payload column, at every shared graph edge. Construction goes through Bounds::offsets, and the three places here that hand-rolled a Vec<usize> of end offsets by matching both variants now call Bounds::to_vec, which is public and does exactly that.

Value::Sum carries one Tags. The lane assignment — tags plus within-lane offsets — is one type with a two-word Const form for the case where every row carries one tag. signed_order_view moves the assignment through untouched instead of rebuilding it; untranscode reads tag_at/offset_at per row.

One adoption beyond compiling

sort_consolidate was building two i/i+1 index columns to ask corgi for an adjacent comparison — 16 bytes a row of index to produce a one-byte-a-row answer. compare_adjacent names the pattern instead, so the index columns are not built and corgi reads both sides densely. That is on the chunk-merge path.

Verification

• 89 interactive tests pass.
• Clippy is unchanged: 70 findings here, and 70 on a pristine master-next build, so none are introduced.
• Cargo.lock is gitignored, so only Cargo.toml moves.

Note for review

Porting this is what caught a real gap in the upstream PR: Tags was exported but every reader was pub(crate), so the type could be named and not used — which untranscode needs. That went back as an extra commit on WIP#18 before it merged, so the pin here is the fixed version. Worth doing the downstream port before merging the upstream change rather than after.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PPtSwqXTuH4wefbQd2F8pA