Design 47 H: H-A-342, a tool runner's restart kills calls in flight

This commit is contained in:
2026-11-21 22:08:54 +01:00
parent 70e333055c
commit 3f9240f731
@@ -2106,3 +2106,14 @@ Where two preparers disagreed, the chosen one is named.
and a mover kept or added again is a journal line; the builder's health names an event backlog older than five
minutes; §790's and §803's tests show the clear; dry-run keys deleted at the indexer's start. P-NATS-native,
P-failure-loud, map-onto-how-the-mesh-is-built.
- **H-A-342 (a tool runner's restart kills calls in flight, 22:08)** The delivery fixes' pipeline failed at 21:59: the
control-node's node-engine restarted its tool runner for a changed fingerprint in one module's words while the
forge's bundle was taking a package's first chunk, so the publish died (SIGTERM) and the contracts module's build
failed; every mover downstream was not asked. The next merge's pipeline carried the same movers and the build
passed on its second try. Two lines for H2: (1) a publish or fetch refused by a holder away (no responders, a
bundle exited) is a transport fault: the build module ends the run again (§799 covers publish as fetch), the
runner's delays carry the retry; (2) the tool runner drains at a stop: it takes no new call once stopped, finishes
the calls in flight within a bound (setting drain-bound, 30 s), then exits, and the node-engine waits that bound
before it starts the new one; a bundle's call that outlives the bound is cut and said. No repository fix for this
instance; nothing to rerun (main.155 carries it). P-NATS-native, P-failure-loud, fix-a-bug-met-twice (the restart
loop of row 179 killed calls the same way).