perf: binary-search index leaf position instead of full-page decode

[iheitlager]

Summary

• insert_into_index_leaf and delete_from_leaf each fully decoded every cell
  on an index leaf page (collect_index_leaf_cells) just to linear-scan for
  one insert/delete position, and decode_value_cell re-copied the entire
  page into a fresh Rc<[u8]> once per cell decoded rather than once per page.
• Flamegraph profiling of tests/performance/crud.rs (triggered by
  benchmark results showing all write ops 5-11x slower than the SQLite
  oracle, with update_filtered_range at 85x) traced the shared bottleneck
  to this exact code path.

Changes

Changed

• src/btree/index.rs: decode_value_cell now takes a shared Rc<[u8]>
  built once per page by its callers instead of building its own per cell.
  Added search_index_leaf — a binary search over an index leaf's sorted
  cells, decoding only the O(log n) cells actually compared — and
  value_cell_len, a cheap byte-length-only check (no key decode) for
  space bookkeeping.
• src/btree/index/insert.rs: insert_into_index_leaf uses
  search_index_leaf for position/duplicate-check and a cheap length-only
  loop for the page-space check. Full collect_index_leaf_cells now only
  runs on the already-rare split/fragmented-fallback paths, which genuinely
  need every cell's contents.
• src/btree/index/delete.rs: delete_from_leaf uses the same
  search_index_leaf swap — splice_delete_cell already recomputes the
  cell span from raw bytes by index, so only the matched cell's bytes (for
  overflow_page_of) are needed.

Not touched (intentionally out of scope, see #648 discussion)

• IndexCursor::seek's read-path linear scan — a documented Tier 0 design
  tradeoff (index.rs:20-22), a separate concern.
• Interior-page collection (collect_index_interior_entries, used in
  split-promotion/predecessor-swap) — gets the Rc-hoist fix but not a
  binary-search rewrite, since those call sites already need full page
  contents.

Testing

• cargo test --lib — 959 passed, 0 failed
• make test (unit, public-API, proptest, doctests) — all green
• make lint (clippy + fmt) — clean
• make -C tests/performance crud (criterion, sqlite-rs vs pinned oracle):
  • update_filtered_range (the reported 85x outlier): 666ms → 123ms (5.4x faster)
  • delete_pk: 8.7% faster; delete_filtered_range: 2.7% faster
  • insert_single, update_indexed_column, delete_equality_bucket: no
    change (within run-to-run noise) — expected, since single-row writes
    are dominated by fixed per-statement fsync/journal cost, not index
    decoding
  • No regressions found across the 15-scenario suite

spend: matched estimate (medium)

Related Issues

Closes #648

---

🤖 Generated with Claude Code