perf(neo4j): use CREATE instead of MERGE for CALLS edges in write_function_call_groups - #1161

…tion scope, and MATCH optimisation

Three compounding issues caused slow CALLS edge writes on Neo4j.

1. Broken backend detection in write_function_call_groups: the legacy
   getattr chain returned the method object (always truthy) instead of
   calling it, silently using MERGE instead of CREATE on neo4j/nornic
   backends — negating PR CodeGraphContext#1161 entirely. Replaced with the canonical
   get_backend_type() call already used elsewhere in the file.

2. Single managed transaction across all CALLS batches (introduced by
   PR CodeGraphContext#1171): execute_write wrapped the entire _work closure — all 100k+
   edges in one transaction. Neo4j held write locks and accumulated WAL
   entries until the final commit, producing 46-minute silent gaps with
   ~167ms/edge observed in production logs. Replaced with a plain session
   so each 1000-edge batch gets its own execute_write — per-batch
   auto-retry and clean rollback without the lock contention.

3. Inefficient MATCH on called nodes: the called MATCH used only
   name+path (2 of 3 composite index properties), causing Neo4j to scan
   all matching nodes and filter line_number in a WHERE clause — 26,680
   DbHits per lookup vs 2 DbHits with the full composite index. Split
   each batch into fast-path (called_line_number known: full 3-property
   composite index hit) and slow-path (ambiguous resolution: 2-property
   prefix scan, correct behaviour preserved).

   Same fix applied to post_resolution.py where the caller MATCH had no
   label (full node scan, no index possible) and the called MATCH had the
   same 2-property issue.

4. Fixed NameError in graph_builder.py where cgcignore_path was
   referenced in _build_graph_from_scip but never passed as a parameter,
   causing SCIP indexing to always fall back to Tree-sitter.

Benchmarks (~97k fn→fn CALLS edges, Neo4j backend):
  Before (MERGE, single tx):           ~138ms/edge, ~1697s total
  After, 1 repo  (~13k functions):       3.3ms/edge,   320s  (43x faster)
  After, 2 repos (~27k functions, CI):   7.1ms/edge,   689s  (19x faster)

Fast/slow path split: 28% fast (known called_line_number, 2 DbHits),
71% slow (ambiguous resolution, index prefix scan). The fast path
delivers 13,340x fewer DbHits per row — proportionally larger benefit
on graphs with hundreds of thousands of nodes.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

Co-authored-by: Cursor <[email protected]>

…tion scope, and MATCH optimisation

Three compounding issues caused slow CALLS edge writes on Neo4j.

1. Broken backend detection in write_function_call_groups: the legacy
   getattr chain returned the method object (always truthy) instead of
   calling it, silently using MERGE instead of CREATE on neo4j/nornic
   backends — negating PR CodeGraphContext#1161 entirely. Replaced with the canonical
   get_backend_type() call already used elsewhere in the file.

2. Single managed transaction across all CALLS batches (introduced by
   PR CodeGraphContext#1171): execute_write wrapped the entire _work closure — all 100k+
   edges in one transaction. Neo4j held write locks and accumulated WAL
   entries until the final commit, producing 46-minute silent gaps with
   ~167ms/edge observed in production logs. Replaced with a plain session
   so each 1000-edge batch gets its own execute_write — per-batch
   auto-retry and clean rollback without the lock contention.

3. Inefficient MATCH on called nodes: the called MATCH used only
   name+path (2 of 3 composite index properties), causing Neo4j to scan
   all matching nodes and filter line_number in a WHERE clause — 26,680
   DbHits per lookup vs 2 DbHits with the full composite index. Split
   each batch into fast-path (called_line_number known: full 3-property
   composite index hit) and slow-path (ambiguous resolution: 2-property
   prefix scan, correct behaviour preserved).

   Same fix applied to post_resolution.py where the caller MATCH had no
   label (full node scan, no index possible) and the called MATCH had the
   same 2-property issue.

4. Fixed NameError in graph_builder.py where cgcignore_path was
   referenced in _build_graph_from_scip but never passed as a parameter,
   causing SCIP indexing to always fall back to Tree-sitter.

Benchmarks (~97k fn→fn CALLS edges, Neo4j backend):
  Before (MERGE, single tx):           ~138ms/edge, ~1697s total
  After, 1 repo  (~13k functions):       3.3ms/edge,   320s  (43x faster)
  After, 2 repos (~27k functions, CI):   7.1ms/edge,   689s  (19x faster)

Fast/slow path split: 28% fast (known called_line_number, 2 DbHits),
71% slow (ambiguous resolution, index prefix scan). The fast path
delivers 13,340x fewer DbHits per row — proportionally larger benefit
on graphs with hundreds of thousands of nodes.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>