feat(cypher): implement shortestPath execution with igraph/cugraph backends #1010

Summary

This is the active direct Cypher shortestPath(...) execution lane.

Active implementation PR:

• #1042 feat(cypher): execute shortestPath scalar queries #1042

Current branch status on April 3, 2026

#1042 currently adds the narrow scalar shortest-path contract needed by the benchmark lane:

• length(path)
• path IS NULL
• CASE path IS NULL WHEN true THEN -1 ELSE length(path) END
• exact official comma-seeded interactive-complex-13 shape
• bounded, reverse-direction, disconnected-null, endpoint/id-projection, staged WITH, zero-hop, cyclic/bounded duplicate-route, and multi-row shortestPath amplification around that scalar slice

Still intentionally unsupported on the branch:

• allShortestPaths(...)
• RETURN path
• relationships(path) and other generic path-carrier forms

Focused validation so far

Local focused checks:

• graphistry/tests/compute/gfql/cypher/test_parser.py -k shortest_path
• result: 1 passed, 95 deselected
• graphistry/tests/compute/gfql/cypher/test_lowering.py -k "shortest_path or allShortestPaths"
• result: 40 passed, 660 deselected

Focused dgx-spark host-CPU checks:

• parser: 1 passed, 95 deselected
• lowering/execution: 40 passed, 660 deselected

What this branch found while implementing the lane

Two adjacent runtime fixes were needed in the row-binding path:

• fixed-point edges must count as multihop in graphistry/compute/gfql/same_path/edge_semantics.py
• multihop row bindings must preserve hop-labeled node frames into the next node filter in graphistry/compute/gfql/row/pipeline.py

Seven adjacent scalar/runtime hardening fixes were then needed during amplification:

• shortest-path scalar rewrites now bind onto the endpoint alias hop property, and disconnected empty-row synthesis now carries alias-prefixed seed values so endpoint/id projections survive direct RETURN null-extension
• disconnected shortest-path rows now also null-extend through staged scalar WITH ... RETURN ... chains in the supported local subset
• unbounded shortestPath self-matches now produce valid zero-hop scalar results instead of null / -1
• cyclic shortestPath scalar execution now dedupes equal-length revisits instead of erroring or diverging on multiroute graphs
• bounded cyclic and self-loop shortestPath scalar outputs now dedupe identical shortest rows instead of returning duplicates
• multi-row shortestPath pair queries now preserve row-specific endpoint hop values instead of collapsing to 0 / 1
• multi-row disconnected shortestPath now null-extends per seed pair for direct, staged, bounded, and reverse-direction scalar forms

Recent test-only amplification on top locks:

• multi-row path IS NULL ordering across direct, staged, bounded, and reverse-direction shortestPath queries

So this issue is well past parser recognition: the active branch is executing the shortest-path scalar contract end-to-end in focused tests and has gone through several real amplification passes that kept uncovering product bugs.

Benchmark target

Official query file:

• /home/lmeyerov/Work/pyg-bench/third_party/ldbc_snb_interactive_v1_impls/cypher/queries/interactive-complex-13.cypher

Official query text:

MATCH
    (person1:Person {id: $person1Id}),
    (person2:Person {id: $person2Id}),
    path = shortestPath((person1)-[:KNOWS*]-(person2))
RETURN
    CASE path IS NULL
        WHEN true THEN -1
        ELSE length(path)
    END AS shortestPathLength