feat(cypher): support variable-length rels in connected patterns (#973)

[lmeyerov]

Summary

Support variable-length relationships in connected multi-relationship patterns (#973). Previously rejected, now MATCH (a)-[:R*2]->()-[:R]->(c) RETURN c works.

Before

MATCH (a)-[:LIKES*2]->()-[:LIKES]->(c) RETURN c.name
→ GFQLValidationError: only supported as the only relationship in a connected pattern

After

g.gfql("MATCH (a)-[:LIKES*2]->()-[:LIKES]->(c) RETURN c.name")  # exact hops
g.gfql("MATCH (a)-[:R*1..3]->()-[:R]->(c) RETURN c.id")          # range hops
g.gfql("MATCH (a)-[:R]->()-[:R*2]->()-[:R]->(c) RETURN c.id")    # middle position
g.gfql("MATCH (e)<-[:R*2]-()<-[:R]-(c) RETURN c.id")             # reverse

Implementation

Add prune_to_endpoints flag on ASTEdge that filters the graph to only the variable-length hop's endpoint wavefront after traversal. Uses edge hop labels to identify max-distance edges, then prunes nodes + edges to only those endpoints. The lowering sets this flag on non-terminal variable-length relationships in connected patterns.

Tests (11 new)

• Exact hops at start (*1, *2, *3 then fixed)
• Range hops (*1..2, *1..3 then fixed)
• Fixed then varlen
• Varlen in middle (fixed-*2-fixed)
• Branching graph (star topology)
• Reverse direction
• No match (empty result)
• Open range (*0.., *1.. then fixed)

GPU validation

All pass on pandas + cudf (dgx-spark, graphistry/test-gpu:latest).

Test plan

• 450 lowering tests (11 new, 0 regressions)
• GPU: pandas + cudf on dgx-spark
• Adversarial: 14/15 cases pass (1 pre-existing empty graph issue)

🤖 Generated with Claude Code