perf/fix: V3 CALLS batching, watcher fixes, incremental rebuild, delete_repository robustness - #796

…ectory hierarchy

The index_helper 'already indexed' check used a single-hop query
(Repository)-[:CONTAINS]->(File) which misses files behind directory
nodes in the hierarchical graph structure. Changed to [:CONTAINS*]
to traverse Repository->Directory->...->File chains correctly.

…ion on large repos

…umulation on watch events

…ad of O(n)

Previously every file change event triggered:
 - full re-scan of all repo files for imports_map (~28k files for nucleus)
 - full re-parse of all files into all_file_data
 - wipe and recreate ALL CALLS/INHERITS for the entire repo

New approach (O(k) where k = changed file + its callers):
 - query Neo4j BEFORE deleting nodes to find which files call into the changed file
 - update imports_map for the changed file only (O(1) file scan)
 - update_file_in_graph: DETACH DELETE automatically cleans up all incoming/outgoing
   CALLS and INHERITS on the changed file's nodes
 - delete outgoing CALLS only from affected caller files (not the whole repo)
 - re-parse only the affected subset (changed file + callers + inheritors)
 - build file_class_lookup from Neo4j query — no full re-parse needed
 - re-create CALLS/INHERITS for the subset only

New graph_builder.py methods:
  get_caller_file_paths(file_path_str)
  get_inheritance_neighbor_paths(file_path_str)
  delete_outgoing_calls_from_files(file_paths)
  delete_inherits_for_files(file_paths)
  get_repo_class_lookup(repo_path)

_create_all_function_calls gains optional file_class_lookup parameter.
watcher.py gains _update_imports_map_for_file(changed_path).

Tested against local neo4j-cgc — all assertions pass.

…er file

Previously add_file_to_graph made N individual session.run() calls per file:
 - 1 per function node
 - 1 per parameter (×avg params per function)
 - 1 per class node
 - 1 per variable node
 - 1 per import statement
 - 1 per class→function CONTAINS link
 - 1 per nested function CONTAINS link
 - 1 per directory level in the hierarchy
For a Java file with 10 methods × 3 params + 5 imports = ~45 round-trips per file.
For nucleus (~28,600 files) this was over 1 million individual Neo4j transactions.

New approach: one UNWIND batch per node type per file.
 - Functions, Classes, Variables, Interfaces, Macros, Structs, Enums, etc.:
   one UNWIND $batch query per label (replaces N individual MERGEs)
 - Parameters: one UNWIND batch for all function params in the file
 - Class→Function CONTAINS: one UNWIND batch
 - Nested Function CONTAINS: one UNWIND batch
 - Imports (JS): one UNWIND batch
 - Imports (other): one UNWIND batch
 - Ruby modules: one UNWIND batch
 - Ruby module_inclusions: one UNWIND batch

Also removed:
 - Stray info_logger("Executing add_file_to_graph with my change!") on every file
 - info_logger("Processing import: ...") on every import statement
 - Pre-seeded repo_path_str parameter: callers can supply the resolved repo path
   string to skip a DB round-trip (MATCH Repository) per file — used by the
   build_graph_from_path_async loop which knows the repo path upfront.
 - asyncio.sleep(0.01) per file → sleep(0) every 50 files.
   Old: 28,600 × 10ms = ~286s (~5min) of pure dead wait for nucleus.
   New: ~0s overhead (yields to event loop only when needed).

Smoke-tested against local neo4j-cgc (admin123) — all node types and
relationship types verified correct after the rewrite.

CONTAINS* chain delete only removes nodes reachable from the Repository
node. Nodes orphaned by a previous crash or partial index survive and
retain their old CALLS/INHERITS relationships. When the next --force
reindex runs MERGE to recreate nodes, it matches these orphaned nodes
and stacks new CALLS on top of the old ones, producing duplicates
(755k → 1.05M on nucleus).

Phase 2 batches a path-prefix MATCH+DETACH DELETE (5000 nodes at a time)
to catch any orphaned nodes not in the CONTAINS chain, ensuring the graph
is truly empty before the new index is written.

DETACH DELETE on nodes with many relationships causes Neo4j to exceed
transaction memory limits (11.2 GiB on remote:
Neo.TransientError.General.MemoryPoolOutOfMemoryError).

Fix: in Phase 2, delete CALLS/INHERITS/CONTAINS relationships in batches
of 10000 first (both outbound and inbound), then delete the now-bare
nodes in batches of 5000 using plain DELETE (no cascade needed).

…tesian explosion

…ck indexed check