feat(gfql/cypher): EXISTS { pattern } subqueries — declarative prune-isolated, native on 4 engines

[lmeyerov]

Implements EXISTS { <pattern> } pattern-existence subqueries (openCypher-standard) in WHERE position, native on all four engines (pandas/cuDF/polars/polars-gpu) — the declarative prune-isolated building block for the streamgl-viz filter pipeline (frozen semantics: the panel-algebra doc, viz-filter L1). Stacked on #1677.

What's added

• Parser: EXISTS { <pattern> } / NOT EXISTS { ... } in WHERE. The body parses into the SAME BooleanExpr(op="pattern") leaf as a bare pattern predicate — zero new lowering machinery; NOT composes through the existing lift into anti_semi_apply. Inline block comments and property maps ({w: 1}) inside the pattern work. RETURN/WITH-position EXISTS and the not((a)--(b)) spellings keep their clean fail-fast.
• Existential locals: aliases introduced inside the braces are allowed (EXISTS { (n)--(m) } with m unbound outside) — the correlated bindings project them away. Bare pattern predicates keep the conservative no-new-aliases guard.
• Drop-self flavor: the one supported inner WHERE form is endpoint inequality — EXISTS { (n)--(m) WHERE m <> n } ("has a neighbor other than itself"). Carried as a safelisted neq= param on the semi-apply ops: pandas/cuDF filter the correlated bindings table; polars excludes self-loop edges (exactly the m <> n witness for a single-edge pattern).
• Polars goes native (was honest-NIE): semi_apply_mark / anti_semi_apply (correlated key sets via the polars chain executor's named-flag columns — the backward prune makes the flag exactly "a full pattern match exists through this node"; order-preserving is_in instead of joins, null keys matching pandas merge semantics) and rows(binding_ops=...) for the single-entity row table (mirrors the pandas alias-lookup-frame layout). v1 declines honestly: multi-alias correlation, non-single-hop shapes, query= nodes.

Prune-isolated, declaratively (the point)

MATCH (n) WHERE EXISTS { (n)--() } RETURN n              // keep-self flavor
MATCH (n) WHERE EXISTS { (n)--(m) WHERE m <> n } RETURN n // viz drop-self flavor

Both flavors are oracle-pinned in the conformance matrix on a discriminator graph (an isolated node + a self-loop-only node), 4-engine parity-or-NIE.

Tests / verification (dgx GB10, --gpus all)

Full polars lane + conformance matrix + GPU suite + row-pipeline ops = 2082 passed, 0 failed (final tip); cypher lowering suite 1193+ passed (EXISTS marker-shape parity with bare patterns, NOT→anti, neq threading, boundary declines). End-to-end sanity: NOT EXISTS {(n)--()} returns exactly the isolated node on pandas == cuDF == polars. Conformance-ledger waiver text updated to native status (join_apply remains the honest-NIE waiver). ruff + mypy clean.

Honest boundaries (clear errors, no wrong answers)

EXISTS in RETURN/WITH projections; general inner WHERE clauses; multi-pattern bodies; full MATCH..RETURN subquery bodies; multi-alias correlation on polars (pandas/cuDF handle it via the pre-existing bindings machinery).

Review hardening (3-wave converged adversarial review, 2026-07-05)

Findings-mode review (artifacts: plans/viz-filter-pipeline/waves/pr1679/) found and fixed, dgx-verified: GRAPH { } + EXISTS was silently dropping the filter (all engines — now fails fast, anchored to the grammar's constructor-at-start rule with zero false positives proven); the working graph-state prune patterns are pinned 4-engine per the graph-shape acceptance requirement (GRAPH { MATCH (a)-[e]-(b) } keep-self with ALL edges; + WHERE a.id <> b.id drop-self); EXISTS-body keyword scans run on masked text (property strings like 'WHERE it hurts' no longer falsely decline); neq validates as an exact string pair (was a late IndexError / silent first-pair filter); self-alias EXISTS { (n)--(n) } declines instead of a non-NIE crash; null-key and alias-collision corners decline honestly; the pandas name-flag column leak is mirrored exactly; native-must/multi-alias-NIE/decline-shape pins added (changed-line coverage gate green). TCK: the openCypher existential-subquery scenario expr-existentialsubquery1-3 is formally promoted to supported in the tck-gfql corpus (mirror branch feat/gfql-cypher-exists); 1-1 tracked as a whole-entity-rendering wrong-rows debt. Waves 2–3 converged at zero BLOCKER/IMPORTANT. Known follow-up (pre-existing at base): bare pattern predicates inside GRAPH { } bodies also silently drop (row_pre_filters unwired in the graph compiler) — L3 item.

Post-convergence tip movements (structure/CI only, re-verified)

pattern_apply.py module split (module-per-family, per review steer): the polars semi-apply/pattern-existence lowerings move out of row_pipeline.py into graphistry/compute/gfql/lazy/engine/polars/pattern_apply.py — dispatch imports updated, no behavior change (dgx 1703-pass; post-rebase import-shape greps). The polars coverage-audit baseline gains the required pattern_apply.py entry (floor 67.0; dgx-measured 75.36% on the CI-verbatim CPU lane, identical on this and the L2 tree) — the audit fails un-baselined target files, which was the only CI red. CI green at the final tip.

🤖 Generated with Claude Code

[lmeyerov]

there is probably fuzzing & tck repo work we should do across the pr's

[lmeyerov]

avoid dynamic typing patterns

[lmeyerov]

avoid dynami typin gpatterns here too