DDIR Corgi join/arrange: four copy removals

[frankmcsherry]

Four small changes that remove a copy or shorten a buffer's lifetime in the Corgi join and arrangement path. There is one commit each; each can be reviewed or dropped on its own.

1. Borrow join comparison lanes. Comparison views used to gather temporary key columns before comparing them. They now borrow the chunk's u64 lanes with absolute row indices. Adds an oracle against structural row comparison.
2. Skip the discarded hash column. Join output used to gather a compound key's identifier (hash) column and then drop it. It now borrows the declared key instead.
3. The chunker takes the first input's buffers. When the accumulator is empty, the first input's time and diff buffers are swapped in instead of copied, and the chunker's empty buffers go back to the sender. Adds an oracle.
4. Release merge planning buffers at their last use. The survey and offset buffers are dropped before the merge output is materialised, instead of at the end of the merge.

2 files, +135 −65.

The case for this is less copying and less held memory, not speed. Base is master-next 5fd3ecc and the candidate is this branch. Three pairs, order reversed in the middle pair, on an M4. The tour and SCC programs have their inspects removed. Medians:

Workload	Workers	Initial tick	Ten churn ticks	Peak footprint
tour 5k/10k	1 / 4	+0.4% / +0.4%	+0.1% / 0.0%	0.0% / +0.8%
tour 50k/100k	1	−0.4%	−0.5%	−2.4% (0/3 higher; ranges separated)
tour 50k/100k	4	+1.1%	−0.1%	−0.5%
SCC	1	−1.1% (0/3 slower)	+0.1%	−1.3%
SCC	4	−1.4% (0/3 slower)	−0.2%	see below
AoC runner (wall)	1 / 4	−1.4% / −0.5%

• Time: within ±1.4% everywhere.
• Memory: the large tour's 1-worker peak falls 2.4% in every pair.
• SCC 4-worker peak is noise. The first run showed +7.4%. A second run tested each cumulative prefix against base. Base alone ranged from 78.5 to 93.1 MiB over 12 runs (about ±9%), and no prefix separated from base. The full branch came out +5.8%, higher in 2 of 3 pairs, inside that spread. Of the four commits, only the chunker change (3) could plausibly hold more memory: it keeps the donated buffers' full capacity. If that matters, dropping commit 3 is independent of the others.

Earlier measurements on the worldgen DDIR tour:

• the chunker change cut cumulative large allocations by 2 to 3%;
• the merge release lowered the live-heap high-water from 787 to 759 MiB (−3.6%);
• the hash-column skip removed about 26 MB of discarded copies per tour;
• time was flat in all three.

Tests. cargo test --release -p interactive: 114 passed, 5 ignored (all ignored before this change too). All 33 AoC answers are checked on every run, and every timed tour and SCC output matches an oracle agreed by both backends.

The changes are by Astra (OpenAI Codex). I cherry-picked them onto master-next, resolved a test-only conflict with #908, and ran these measurements.

🤖 Generated with Claude Code