v5.10.4

RDR-151 root-cause fix for the recurring T2 daemon 100% CPU peg, plus a T1 scratch
recovery fix.

Fixed

• T2 daemon CPU peg root-caused and fixed (RDR-151, nexus-uzay8 / nexus-xmohw).
  The peg was contention-driven write starvation: taxonomy.* write RPCs bypassed
  the daemon's _catalog_write_lock, and nx index opened its own direct T2 write
  connection, so multiple writers raced the single SQLite WAL writer lock while the
  daemon's executor thread wedged on it. Two fixes: (1) taxonomy.* writes now
  serialize through the daemon's catalog write lock; (2) every direct T2 taxonomy
  write on the nx index path and the nx taxonomy CLI now routes through the
  daemon (compute client-side where ChromaDB is needed, persist daemon-routed),
  leaving only read-only direct connections (WAL readers don't take the writer
  lock). A storage_boundary_lint enforce-flip fails CI on any taxonomy write
  method called on a directly-constructed handle outside db/+daemon/.
• T1 scratch restored when the session id diverges from the MCP lease key
  (nexus-gff3g). T1 working memory is now matched by Claude-ancestor pid with a
  deterministic tie-break, so sibling agents recover the correct scratch when the
  session id and lease key differ.
• Daemon lifecycle/teardown hardening: idle read deadline on accepted connections,
  mark_shutting_down() first in stop() (T2 + T3), to_thread(heartbeat_tick)
  so the flock never blocks the event loop, and reaper endpoint.pid fallback.

Changed

• RDR-147 (multi-hop entity-resolution retrieval) accepted.