v4.2.3

v4.2.3 (2026-10-11)

Fixed

• /api/cypher (cypher_api) answered status OK, rows [] for a statement that failed (Mindwalk:
  MATCH (n) RETURN count(n) AS c logged <PARAMETER ERROR> Invalid number of parameters and returned an
  empty result). A failed statement is now a 400 with the error. execute_sql also sends a statement that
  has no placeholders with no parameter list, which is the shape the driver rejected. The reproduction of
  that error needs their populated graph; the error surfacing is tested, the parameter shape is the likely
  cause and is not proven.
• A failed edge-vector step no longer poisons its resume: an empty kg_EdgeEmbeddings_ivg400 or
  edge_vector_quarantine left at a different width is rebuilt, and one that holds rows at the wrong width is
  refused with a message that names the width.

Added

• upgrade_to_4_0_0 logs when each step starts and finishes, the rows placed so far (every 25 batches), and a
  line before the vector-index build, which is the long silent stretch (about 20 minutes for 206K 768-dim
  vectors): "Building the vector index on kg_NodeEmbeddings; this can take many minutes".

Known

• Variable-length Cypher patterns ([*1..2]) read the ^KG adjacency index, so edges written with SQL
  (store_edge, a loader) return nothing until engine.sync(); create_edge writes both. Single-hop patterns
  read SQL and do not need it.
• A database created by older hand-written DDL keeps a rdf_edges without edge_id after
  initialize_schema and upgrade_to_4_0_0 (they only create absent tables), which breaks undirected Cypher
  patterns. Repair it before upgrading (stage the rows, drop the SQLUser.rdf_edges compatibility view, drop
  and recreate the table, restore with a key-level dedupe on MIN(%ID)).