A node's plans took 2.6-13.6 s of CPU on the provisioner: one composition
resolved every machine of the mesh some two hundred times (who is on the
private network, and the rest of the mesh, worked out again on every
look), three compositions did so again each, and the holder composed
twice. A composer now works each out once over its unchanging Input;
Composing gives Resolve, Render and Generators one composer;
PlansHeldTwice composes the wanted and the newest builds over one read of
the bus, says where the time went by phase, and the states reader keeps a
build's take-in by its content. WatchedStates takes a process's own
writes into the copy at once, so a holder reads its own states from it.
Measured on a captured novox mesh: 11.7 s to 0.49 s per composition of an
unchanged node, plans and provisions identical on all four nodes.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
mesh/check pass 790d1c82d027 on da41823855a9 | unit 1964 run, 5697 kept, 30 skipped | integration 1999 run, 5829 kept, 38 skipped | over bound: unit, …
A read could start a state's watch before its follower joined, and a change landing in that gap woke nothing (review
of ADR 0580 §2): a joining follower now runs over the copy's current values before any later change. The watched
reader keeps each archive once per URL, and both caches are bounded.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
A whole state read key by key took the provisioner's composition of one node past 30 s and the assigner's compare past
its call bound (ADR 0580 clause 1): each state is now one watch's current values, kept by its changes; a build is read
by key and a manifest once per digest.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL