v0.10.0

Added

• Database::prepare(cypher) -> PreparedPlan: a genuine prepared
  statement that parses and validates once. Execute it any number of
  times with Database::execute_prepared_plan(&prepared, &params, &options) or the Transaction::execute_prepared_plan twin; repeat
  calls skip re-validation. The existing execute_prepared_statement
  path (pre-parsed text, re-validated per call) is unchanged (#216).

Changed

• Label checks on an expansion's far endpoint ((a)-->(b:Person)) read
  the node's label ids directly instead of decoding its whole property
  map and throwing it away. Measured on a 10K-node, 5-edges-per-node
  graph: label-filtered 1-hop with count(*) 18.1 µs → 13.9 µs
  (−23.5%), with a property projection 17.8 µs → 16.1 µs (−9.6%) (#214).
• The LogicalPlan-driven read path (scans, Expand, Filter)
  resolves Cypher variables to fixed row-slot indices at plan-build
  time instead of hashing a HashMap<String, Binding> per row. Slot
  compilation still runs once per execution, so very small result sets
  pay a few hundred nanoseconds more up front; larger expansions
  amortize it (#215).
• Dependencies: redb 4.2 → 4.4, antlr4rust 0.5 → 0.6 (parser
  regenerated), rustls → 0.23.46 (RUSTSEC-2026-0285/0346/0348/0349),
  and, for the arrow feature, arrow-array/schema/buffer 59 → 60, so
  Database::query_arrow callers must be on arrow 60 (#226).
• Book rewritten around an end-user journey (Getting Started, The
  Cypher Guide, Embedding, Reference, Operations), with Internals
  demoted to an appendix; stale roadmap items and essay-length comments
  trimmed across the workspace (#211, #212, #213).
• RETURN DISTINCT ... LIMIT k with no ORDER BY and no aggregation now
  stops the read pipeline as soon as k (plus SKIP) distinct projected
  rows exist, and variable-length traversals enumerate their paths
  lazily per input row instead of materializing every path up front —
  together, a [:KNOWS*1..3] ... RETURN DISTINCT ... LIMIT 20 over a
  high-degree graph walks a few hundred paths instead of ~1M (measured
  643 ms → 1.84 ms, and 71 ms → 714 µs on a fixed-2-hop DISTINCT+LIMIT,
  LDBC-style SF 0.1). Emission order within a var-length hop is now
  depth-first (row order without ORDER BY is unspecified), and a *0..
  traversal that would exceed the safety depth cap still errors on any
  fully-consumed query, but may now succeed if a satisfied
  DISTINCT+LIMIT stops the enumeration before reaching the offending
  branch.

Fixed

• Start-point selection prices unindexed literal-equality filters
  correctly: the endpoint's emitted-row estimate gets a default 1/10
  equality selectivity (its physical scan cost stays full), instead of
  the filter only doubling that endpoint's cost and inverting the
  anchor choice. Measured on the LDBC-style SF 0.1 workload without
  indexes: 33.9ms → 5.6ms (point lookup + hop) and 650ms → 9.6ms (the
  unlabeled-far-endpoint shape that anchored an AllNodesScan) (#208).

Install: cargo install marsdb-cli, pip install marsdb, or brew install knoguchi/marsdb/marsdb.