0.59.0 — GFQL correctness, native Polars, and verified benchmarks

The headline: 0.59.0 makes absent-schema behavior explicit, closes a broad set
of silent-answer and cross-engine correctness gaps, expands native Polars and
resident-index execution, and replaces untraceable benchmark claims with
contract-checked measurements.

Breaking behavior

• GFQL absent labels and properties now use one strictness contract across
  validation, local execution, and remote execution: "strict" raises,
  "warn" (the new default) warns once per absent name and resolves it to null,
  and "quiet" has the same openCypher answer without warnings. Legacy
  strict=True / strict=False remain aliases for "strict" / "quiet".

Correctness and safety

• Fixed silent multiplicity errors in whole-entity returns, seeded hops, and
  OPTIONAL MATCH; aligned endpoint closure, null endpoints, variable-length
  paths, aggregate types/errors, temporal expressions, and index invalidation
  across engines.
• Remote GFQL/Python now preserves typed error surfaces, rejects unsupported
  engines before credentials/uploads/requests, preserves None filters and
  output parameters, and refuses lossy CSV decoding instead of silently
  rewriting values.
• Graph-preserving Cypher CALL returns the requested Polars engine instead of
  leaking pandas frames. Explicit Polars-GPU paths no longer report CPU fast
  paths as GPU execution.

Native Polars, indexing, and performance

• engine="auto" follows bound Polars frames into the native Polars engine and
  can target Polars GPU for genuinely usable all-cuDF inputs.
• Added native directed unbounded variable-length bindings, secondary node
  property indexes, verified column-stat and degree facts, index-planner
  diagnostics, and fast-path engagement in gfql_explain.
• Reduced repeated compilation, eager materialization, redundant semi-joins,
  wide projections, and degree-table work in grouped aggregates and two-hop
  count(*), with proof-gated declines to the established general paths.

Benchmarks and documentation

• The GFQL q1–q9 board and filter → PageRank → filter page now render from a
  vendored pyg-bench artifact plus contract v3. CI rejects stale, diagnostic,
  disclosure-dropping, or unproven derived numbers.
• The completed 30,494,866-edge GPlus Neo4j + GDS arm reports a direct 354.47 s
  median-of-slot-medians with exact selected-node parity. No GFQL-vs-Neo4j
  ratio is published because the profiles differ: Neo4j includes server round
  trips and rebuilds its GDS projection per iteration, while GFQL retains
  resident frames.
• Removed benchmark figures without surviving raw measurements, commits,
  environments, or reproducible derivations; restored reproducible charts and
  added/updated GFQL strictness, engine, indexing, performance, and benchmark
  guidance.

Release assurance and follow-ups

• Exact-head CI covers Python 3.8–3.14, GFQL/TCK, Polars 3.9–3.14, compatibility,
  AI/UMAP, Spark, Neo4j, package build, docs, benchmark contracts, changed-line
  coverage, actionlint/zizmor, and Read the Docs.
• GPU claims remain bounded by the repository's documented lack of a public
  cuDF CI lane; no new unsupported GPU-wide claim is made here.
• Non-blocking maintenance follow-ups: Spark Python 3.14 security-pin refresh
  #2014 and Node
  24-native GitHub Actions upgrades
  #2015.

Links

• Full 0.59.0 changelog
• PyPI 0.59.0
• Read the Docs 0.59.0
• Latest documentation