Design 47 G: the lead's answers A-487 and A-487a (the catalog answers from memory; a node-engine asks a missing trunk build again on a backoff) #717

Merged
mesh-tools merged 1 commits from design/47-G-lead-answers-126 into main 2026-11-18 23:54:00 +00:00
@@ -1977,3 +1977,16 @@ the principle that settles it, cited as "design 47 G A-n".
answer, refusing before it asks only what the verb itself refuses, as `runner_rerun` does; the hand act is recorded
by the verb. The seat has no verb `release`, so no tool is made for it. (From the prepare of the take-in of the
modules each node runs at builds placed by hand before G: a cancelled pipeline could not be started again.)
- **A-487** (G §421; principles 1, 3 and 7) The catalog half answers every verb of the seat `mesh-catalog` from the view
of the newest trunk revision it holds in memory, with no call to the bus per read: it reads the newest revision once
when its runtime starts, and again only when it hears the seat's own event `catalog-revision` naming a newer trunk
revision; it never lists its objects `catalog-revisions` to answer a read. The verb `module` reads the one module it
is asked, never every module. Each read of a new revision is said with its revision, its module count and how long
it took (principle 3). (From the recovery of 2026-11-19: every key read listed the whole bucket, so one `module`
answer cost about 150 listings, and the node-engines' asks starved every other catalog call; the second of the kind
after the kept view of the trunk.)
- **A-487a** (A-461e, A-487; principles 1 and 7) A node-engine asks a module's trunk build of the seat `mesh-catalog`
again after an answer that was none only when the catalog says `catalog-revision`, or after a backoff of 5 s, 15 s,
30 s, 1, 2 and then 5 minutes per module, whichever comes first; never on every pass. The backoff of a module starts
again from 5 s when an answer comes. (Same recovery: three asks at a time per node, asked again at every pass,
kept the catalog choked after each restart.)