GeezerVison

Before: ResultSlicer accumulated provider batches to fill the requested chunk (100,000 rows by default), removed the needed prefix with splice(), retained the tail, and joined pieces with flat(). Repeated fetches shifted the leftover rows; fetchAll() also used flat().

After: Keep the batch and advance an offset, taking requested rows with slice(). Compact the tail after more than half is consumed to release row references without repeatedly copying a large tail. Reuse a single piece directly, or join multiple pieces with a copy loop shared by fetchAll(); cap initial allocation to avoid V8 dictionary arrays. Chunk sizes, order, and buffering behavior stay the same.

Validation: 1,548 unit tests, lint, formatting, type checking, and build pass. Offline 80M-row Arrow pipeline: 29.7 → 13.8 ns/row.

Live kernel benchmark on pecotesting (Medium serverless), result caching and row materialization disabled, 100,000-row chunks. Medians of five fresh-process samples per build after warmup; all row counts matched.