perf(neo4j): fix O(N²) edge inserts via labeled MATCH + UNWIND batching

[kingjin94]

Problem

push_to_neo4j was catastrophically slow on large graphs (5+ hours for ~40k nodes/edges in production). Two compounding issues:

1. Per-row round-trips: each node/edge was a separate session.run() auto-commit transaction (~77k Bolt round-trips for a medium repo).

2. Label-less edge MATCH (the critical one): the edge query was:

   MATCH (a {id: row.src}), (b {id: row.tgt})

   With no label, Neo4j ignores all uniqueness constraints and scans the full node store for every row — O(N²) regardless of indexes or UNWIND batching.

Fix

1. Group nodes by label and edges by (src_label, tgt_label, rel), send as UNWIND batches of 500 via execute_write().
2. Include labels in the edge MATCH by building an id→label map during the node pass:

   MATCH (a:Code {id: row.src}), (b:Document {id: row.tgt})

   This hits the per-label b-tree index (requires a uniqueness constraint on id per label, which callers should create before pushing).
3. Log progress every 60s so long pushes aren't silent.

Benchmark

Ephemeral Neo4j 5, no pre-existing data:

Strategy	10k nodes / 20k edges
baseline (per-row session.run)	88s
UNWIND + constraint, label-less MATCH	71s (barely better)
UNWIND + constraint + labeled MATCH	1.5s (50x faster)

Production: 212k nodes + 398k edges pushed in ~24s.

Notes

• The label-less MATCH fix is the dominant improvement; UNWIND batching alone barely moves the needle because it just reduces transaction overhead while the per-row full-scan cost remains.
• Callers should CREATE CONSTRAINT label_id_unique IF NOT EXISTS FOR (n:Label) REQUIRE n.id IS UNIQUE for each label before pushing to get the full benefit.