fix(gfql/cypher): neo4j round() tie-breaking + lower/upper GQL aliases (#1673)

[lmeyerov]

Stacked on #1675. Result of a deep standards re-vet of the #1673 dialect extensions against primary sources — the current neo4j Cypher manual, neo4j's shipped Cypher 5/25 parser source, the openCypher TCK, and the ISO-derived GQL grammar + neo4j's GQL-conformance appendix. Full findings table posted on #1673.

Fixes

1. round() tie-breaking — was a wrong answer vs neo4j. We shipped numpy/polars half-to-even (round(2.5) → 2.0). Neo4j's documented semantics: precision 0 (and the 1-arg form) rounds ties toward +∞ (round(-1.5) → -1.0, manual Ex. 8/10); precision > 0 rounds ties away from zero (HALF_UP: round(-1.55, 1) → -1.6, Ex. 11). Fixed on pandas/cuDF (floor(x+0.5) / scaled HALF_UP) and polars (floor(x+0.5) for p=0; native round(mode="half_away_from_zero") for p>0 — a manual scale/divide formula picks up 1-ulp noise from polars' reassociating optimizer). Parity-tested on both engines against the manual's example values. (Note: the neo4j args-table "(default: UP)" contradicts its own Example 10 — docs bug; the examples are authoritative.)

2. lower() / upper() aliases added — ISO GQL mandatory §20.24 character-string functions; neo4j accepts both spellings as GQL-conformance aliases of toLower/toUpper. A GQL-aligned dialect should too.

Spec records (no code change, see #1673 addendum)

• ^ settled as LEFT-associative for the eventual re-add: the openCypher TCK pins left (4^(3*2)^3 = (4^6)^3) and neo4j's shipped parser left-folds Pow (Cypher25Parser.g4 + astBinaryFold=foldLeft) — the manual's "Right to left" precedence row is a neo4j docs bug.
• XOR correction: ISO GQL does have XOR (optional GE07; precedence groups with OR).
• GRAPH g = GRAPH { … } and CREATE GFQL INDEX confirmed as documented extensions (GQL §10.1 takes graph references only; GQL has zero index DDL — neo4j's own CREATE INDEX is likewise beyond-GQL).
• LIKE/BETWEEN absence from GQL re-verified hard (grammar's only LIKE is CREATE GRAPH TYPE … LIKE g).

Tests

Neo4j-manual example values asserted on pandas and polars (test_round_neo4j_tie_breaking, test_lower_upper_gql_aliases, updated scalar cases); 1618 tests pass across expr-parser/lowering/row-pipeline/polars suites; ruff + mypy clean. cuDF lane: the fix uses the same engine-portable pattern as floor/ceil (cuDF-native .floor() else np.floor). GPU parity verified on dgx (GB10, --gpus all, 2026-07-05): full polars lane + conformance matrix + GPU suite = 1807 passed, 0 skipped/failed — the hardened round kernel + lower/upper aliases parity-green on cuDF and polars-gpu vs the pandas oracle.

Review hardening (4-wave converged adversarial review, 2026-07-05)

A multi-wave findings-mode review (artifacts: plans/viz-filter-pipeline/waves/pr1677/) found and fixed, dgx-verified: the polars extra's floor corrected to >=1.29 (Expr.round(mode=) shipped in py-1.29.0, not 1.5 — older installs crashed with a raw TypeError); round(x, p>308) is now the identity on BOTH engines (10.0**p overflow guarded on pandas; polars' quantizing [309,325] window + u32 crash above 2^32 guarded on polars); polars round(p>0) normalizes -0.0 like the pandas kernel (pinned by a sign-bit test — value equality can't see it); deterministic tie/hazard matrix cases so exact ties (±1.25 at p=1, 2^51+0.5, the JDK 1-ulp-below-tie case, 1e308, p=400) actually reach cuDF/polars-gpu; the neo4j BigDecimal decimal-string deviation documented (round(2.675, 2) = 2.67 binary-double here vs 2.68 in neo4j); warning hygiene on the guarded overflow paths. Reviewers verified the p=0 kernel ulp-correct and byte-identical cross-engine, and the tie semantics against both the neo4j manual and CypherFunctions.java. Waves 3–4 converged at zero BLOCKER/IMPORTANT.

🤖 Generated with Claude Code