v5.9.3

Catalog moved behind the T2 daemon (RDR-146), closing the GH #1046 starvation.

Fixed

• Interactive nx dt index no longer starved by background indexing
  (#1046). .catalog.db was the last shared-state store still on the
  direct-sqlite3 model, so a foreground catalog write could wait ~30 min
  behind a hook-spawned nx index repo contending on the one SQLite writer
  lock. The T2 daemon now hosts the single rich Catalog (sole .catalog.db
  writer) behind a write-only op whitelist; every consumer write routes
  through it via typed make_catalog_reader / make_catalog_writer factories
  (reads stay local), enforced by the storage-boundary lint.

Changed

• Interactive-vs-batch catalog-write fairness (RDR-146). Inside the single
  daemon, an interactive write (foreground nx dt index / capture /
  incorporate, MCP store_put) opens a short priority window; the background
  indexer polls it and yields over a bounded budget so interactive writes are
  never starved. nx index --on-locked=skip now also defers a yielded catalog
  write to the next idempotent pass; the per-repo advisory lock keeps its
  separate two-same-repo job. New NX_WRITE_PRIORITY=interactive|batch
  overrides the tty-based default.
• Bounded catalog graph traversal. _LinkOps.graph() BFS replaced with a
  single WITH RECURSIVE SQL query (depth cap, cycle detection).