Design 47 H: H-A-302, one composition per machine, module and fingerprint; an older run never overwrites a newer

This commit is contained in:
2026-11-21 20:18:36 +01:00
parent c27f280237
commit a1b26fe7da
@@ -1760,3 +1760,14 @@ Where two preparers disagreed, the chosen one is named.
H-A-296 and H-A-300: a build or test step whose fetch of its source meets the forge away (401 while the store is
down, connection refused, no responders) is asked again with a back-off until its bound, and the step's verdict
names a fetch failure apart from a build failure. P-failure-loud, P3.
- **H-A-302 (one composition per machine, module and fingerprint; an older run never overwrites a newer, 20:18)** Met
live: after the tool runner's own pipeline passed, the home server's composition key for the tool runner at its new
fingerprint alternated between two digests (one resource differed: the runtime process), two runs writing it in
turn, and the node-engine restarted the tool runner at every flip, every 20 to 50 seconds, so the home server's
tools and the mesh MCP server dropped again and again. Owed, H2, and the recovery of the moment is dropping the stale
run with the assigner's own tool: (1) a composition key is written by the newest composition run for that candidate
only (the write carries the run's pipeline sequence, and a lower sequence is refused); (2) a run whose pipeline ended
writes nothing more; (3) the node-engine applies a composition again only when its content differs from what it
runs AND the write is newer than the one it applied, and says in its journal which run's write it applies; (4) the
build seats' steps read tool answers the steps state (it answered empty while steps ran: a tool fault). P3,
P-NATS-native (one owner per key), tools-are-how-a-module-is-spoken-to.