perf(reader): untranspose fastlanes.delta straight into the output window

[dfa1]

Reconciles #343 with #345, both of which fixed #338 independently. I merged #345 without checking for an open PR on the issue first — that's on me, and this PR exists so #343's work isn't lost rather than because #345 needed changing.

Neither implementation dominated the other, so this takes both halves.

From #343

• scatterChunk writes each untransposed value straight to its output index. That removes the chunk-sized untransposed staging buffer fix(reader): decode fastlanes.delta into the arena, not four heap long[] #345 kept, and the separate pass that sliced it. The bounds handling is the nice part: the leading chunk of an offset-sliced array maps to a negative output index and the trailing chunk runs past rowCount, and a single Long.compareUnsigned(at, rowCount) < 0 covers both, because a negative index reads as a huge unsigned value. The stores are a permutation scatter and never vectorize regardless, so the compare costs nothing the untranspose wasn't already paying.
• out.asReadOnly() on the returned segment.
• Element-indexed getAtIndex instead of hand-computed byte offsets.
• The reader-module tests, which build the delta wire form directly, mirroring DeltaEncodingEncoder's transpose-then-per-lane-delta layout. This is the better call: the writer isn't on the reader's test classpath and never emits a non-zero offset, so fix(reader): decode fastlanes.delta into the arena, not four heap long[] #345 had to reach into the integration module to cover offset slicing. These cover it where it belongs, across offsets 1, 7, 1023, 1024, 1025 and 2000 — i.e. every relationship a window can have to a chunk boundary.

Kept from #345

• Window-limited chunk iteration. perf(reader): decode fastlanes.delta without heap staging #343 walked all numChunks and discarded out-of-window stores per element; a one-chunk slice of a thousand-chunk column did a thousand chunks of work. With firstChunk/lastChunk it does one. The two mechanisms compose — the scatter's compare still handles the partial first and last chunk.
• The metadata range guard. perf(reader): decode fastlanes.delta without heap staging #343 removed the heap arrays, which also removed the NegativeArraySizeException path, but left numChunks = (int)(deltasLen / CHUNK) unguarded: a deltas_len of Long.MAX_VALUE drives the loop ~9×10¹⁵ times. That's a hang, which is worse than the OutOfMemoryError it replaced.

Tests

Both suites, ten tests total: #343's wire-form round trips (all eight widths, chunk boundaries, offset slicing, short row counts) plus #345's malformed-metadata cases. The integration round-trip through the real file format stays as-is.

./mvnw verify green across all 17 modules.

Closes #343.

🤖 Generated with Claude Code