What

In the no-score collection path (Weight::for_each_no_score → for_each_docset_buffered),
matched doc ids were handed to the collector's collect_block one
COLLECT_BLOCK_BUFFER_LEN (64) block at a time. For aggregations this is the dominant path
(requires_scoring() == false), and 64 docs per collect_block call under-amortizes the
per-call overhead (virtual dispatch + aggregation bookkeeping).

for_each_docset_buffered now owns a 2048-element heap buffer and fills it through
successive fill_buffer calls over 64-element windows, flushing a single larger block to
collect_block. fill_buffer keeps its COLLECT_BLOCK_BUFFER_LEN (64) window contract, so
no DocSet implementation changes — only how much we accumulate before flushing.

Notes

• The large buffer is allocated with Box::new_zeroed_slice(..).assume_init() to zero directly
  on the heap (no 2048-element stack temporary). This API is stable since 1.92.0, hence the
  MSRV bump 1.86 → 1.92.
• Audited every collect_block implementation: none assume a max slice length of 64 (they
  delegate or iterate the slice; filter wrappers use a growable Vec). Doc comments on
  COLLECT_BLOCK_BUFFER_LEN and SegmentCollector::collect_block updated accordingly — the
  block handed to collect_block may now exceed 64.
• Uses unsafe to reinterpret each 64-element chunks_exact_mut window as &mut [DocId; 64]
  for the fill_buffer call (sound: 2048 is a multiple of 64).

Benchmarks

cargo bench --bench agg_bench on the full (1M-doc) input, single segment, no deletes — so every bench exercises the for_each_no_score → collect_block path. Reported value is the minimum of 3 runs (most stable estimator; the per-run median swung ±5–20% on its own). Apple Silicon (M-series), REUSE_AGG_BENCH_INDEX so both sides hit the identical index.

Summary: median -1.7%, mean -1.6% across 47 aggregations. Negative = faster. Broad improvement on the cheap full-scan aggregations (where per-collect_block overhead dominated); a handful of heavy order_by_card / top_hits benches regress and are within this harness's run-to-run noise.

Tests

cargo build, cargo clippy --all-targets (no new warnings), and
cargo test --lib {aggregation, query::, collector::} all pass.