Cypher WITH ORDER BY / LIMIT not preserved across MATCH re-entry for multi-row results #1342

Summary

WITH ... ORDER BY ... LIMIT k followed by another MATCH is rejected when the prefix produces multi-row results — ordering / limit are not preserved across the MATCH re-entry. This blocks the LDBC SNB IC1 ordered-top-k friend tail.

Repro (minimal, fails on master 41d865ac9)

MATCH (p:Person {id: $pid})-[:KNOWS]-(friend)
WITH friend
ORDER BY friend.firstName ASC
LIMIT 5
MATCH (friend)-[:STUDY_AT]->(uni:University)
RETURN friend.id AS friendId, uni.name AS uniName

Expected vs observed

• Expected: top-5 friends by firstName survive the LIMIT; the trailing MATCH expands STUDY_AT for that subset.
• Observed: GFQLValidationError: [unsupported-cypher-query] Cypher MATCH after WITH does not yet preserve prefix WITH row ordering across MATCH re-entry for multi-row result shapes | field: with.order_by | value: ['friend.firstName']

Source

• Guard: graphistry/compute/gfql/cypher/reentry/runtime.py:340
• Helper: _bounded_reentry_prefix_order_is_safe (same file)
• Asserting tests: graphistry/tests/compute/gfql/cypher/test_lowering.py:9218, :9226

Benchmark evidence

• pyg-bench artifact: results/runs/dgx-spark-ic1-conformance-2026-04-05/probe-results.json
• LDBC SNB lane: interactive-complex-1 / named-shortest-friends
• Official IC1 query has WITH min(length(path)) AS distance, friend ORDER BY distance ASC, friend.lastName ASC, toInteger(friend.id) ASC LIMIT 20 immediately followed by another MATCH (friend)-[:IS_LOCATED_IN]->(friendCity:City).

Acceptance criteria

• The minimal repro above succeeds with the trailing MATCH applied to exactly the top-5 friends ordered by firstName.
• ReentryPlan (or the bounded-reentry runtime) preserves prefix.order_by and prefix.limit through the trailing MATCH for multi-row prefixes.
• Regression tests cover: ORDER BY single column + LIMIT, ORDER BY multi-column + LIMIT, ORDER BY DESC + LIMIT, all followed by re-MATCH that traverses out from the carried alias.

Subsystem / scope

• graphistry/compute/gfql/cypher/reentry/runtime.py (_bounded_reentry_prefix_order_is_safe, the surrounding admit logic)
• graphistry/compute/gfql/cypher/reentry_plan.py (carry-set + ordering metadata)
• Out of scope: ordering across more than one re-entry hop — that is a follow-up.

Related

• Parent umbrella: [FEA] GFQL bindings_table (row bindings) from chain using Yannakakis prune #880
• Architectural follow-up: Cypher/GFQL: introduce a first-class row-carrier IR for multi-stage vectorized row semantics #989 (row-carrier IR — the long-term home for cross-stage ordering)
• Coordination: Meta: benchmark coverage unlock plan for GFQL and GFQL/Cypher #992, Meta: GFQL compiler/planner post-M4 execution DAG (active wave) #1259