Repository navigation
GFQL Cypher: count(*) trail uniqueness + list-param equality leak (round-006 umbrella) #1905
Description
Activity
- added 4 commits that reference this issue
on Aug 15, 2026 - added a commit that references this issue
on Aug 18, 2026 Umbrella verified served on master
e6625ed28— every item is either fixed or split into its own open issue; recommending close after review.- Item 1 (count(*) trail uniqueness): loop fixture rows=count=1; comma-pattern cartesian on the star fixture rows=count=6 with same-edge pairs excluded — both engines, hand-computed oracles.
- Item 2 (list-param
=leak): list-parameter equality now returns 0 rows on both engines (no raw leak). - Item 4 (str-param vs numeric column): pandas answers the conformant 0 rows; polars raw
ComputeErrorleak is gone — now a typed NIE ("numeric-vs-string comparison... no silent fallback; parity-or-error by design") with pandas/cuDF escape hatch. The remaining polars answer-vs-decline asymmetry is the cross-type comparison family tracked in open bug(gfql): cross-type WHERE on native polars leaks bare polars errors (temporal-vs-string, scalar filters) #1880. - Item 3 (absent label): unchanged by design decision pending —
MATCH (n:Nope)still raisesGFQLSchemaError [column-not-found]instead of openCypher 0-rows/null; that owner decision (strict-schema vs conformant default, incl. the no-op env-knob sub-bug) is captured in open design issue GFQL: declared schema + strictness levels for absent labels/properties (design; local vs remote divergence; dead env knob) #1916, so it's served-by-split.
- added a commit that references this issue
on Aug 20, 2026
Round-006 cartesian/params/labels probe (~85 queries × both engines @ d7cf23d; scripts in session scratchpad probe-r6-cart/).
SILENT-WRONG (both engines):
count(*)ignores relationship uniqueness the row lane enforces — one MATCH, two cardinalities again but in the fixed-hop/comma domain: self-loop fixtureMATCH (a)-->(b)-->(c)rows=1 vs count()=2; branching comma(a)-->(b), (a)-->(c)rows=2 vs count()=9 (out-deg² products incl. illegal same-edge pairs); grouped branching count equally wrong.count(a)/sum/collect/count(DISTINCT)/chain-grouped counts are all correct — only the bare-count lane rides the edge product. A guard exists for aggregates needing path multiplicity (typed decline) — count(*) slips around it.$paramin=/property-map runs as isin (WHERE n.v = $l, l=[30,70] → matches members; oracle: scalar=list false → 0 rows). Literal spelling differs too (pandas 0 rows conformant; polars typed error) — param substitution into filter_by_dict semantics leak.Typed-gap deviations: 3. Nonexistent node label = hard GFQLSchemaError column-not-found (incl. OPTIONAL arms, where openCypher says all-null rows); relationships correctly 0-row. Pandas OPTIONAL-arm error leaks binder-internal column name. 4. String-param vs numeric column: pandas typed; polars RAW polars.ComputeError (error-contract divergence).
NIE inventory: comma-cartesian is node-only (disconnected relationship arms decline;
MATCH (a:X) MATCH (b:Y)rejected while comma form works); polars fall-offs (IN $empty/mixed lists, XOR over labels, labels() projection).QUIET inventory: shared-alias joins with correct trail semantics, label conjunction/OR/NOT, params in all positions incl. shadowing, missing-param typed. Fix order: 1 → 2 → 4 (error contract) → 3 (needs semantics decision: strict-schema vs openCypher missing-as-absent — flag for owner).