Design 47 H: H-A-267 to H-A-270, the four faults behind proof 2's failed pipeline

This commit is contained in:
2026-11-21 15:31:15 +01:00
parent 0b6b20c18d
commit 0d489ff1a7
@@ -1535,3 +1535,26 @@ Where two preparers disagreed, the chosen one is named.
watched branch is the one every registered module of the repository holds alike; where they differ or none is
registered, the repository's default branch as the seed reads it from the seat git; a default branch not readable
skips the module, said in the journal, registered at the next kept revision. P2, P4.
- **H-A-267 (a trunk build gets every mesh package it names at its latest published release, 15:31)** Met live: the
Node build module puts only the SDK into a trunk build and leaves every other mesh package a module's package file
names out, so the forge module's build has failed since 2026-11-19 ("cannot find module", the contracts package).
The Node build module places each mesh package a package file names at the latest release the package registry
holds (ADR 0573 §2, as it does for the SDK); a group's or a pipeline's own member builds take precedence where the
fetch names them. The Go build module already does this through the pins. P2 (one rule for every build module).
- **H-A-268 (the bus module's whole test waits for a fill pass, not a first view's answer, 15:31)** The keeper counts
a first view listing no module as none yet and never marks it answered (A-367), so the subtest "the fill pass runs
while the first view is asked" waits for what cannot come; the test waits for the first fill pass and then takes
the first view. In the bus module's integration test. P4 (a test that cannot pass hides every later failure).
- **H-A-269 (the builder hears a published once and quickly; no publish asked again while one is unheard, 15:31)**
Met live: the builder's consumer of the registry's published events stands thousands behind (each event read against
every reservation, fifteen seconds apiece) while its publish-again loop asks the same release again every few
minutes, adding events. The builder finds a published event's reservation by its digest in one read (an index by
digest kept beside the reservations), and asks a publish again only when no published for that release is unheard
(its consumer's pending below the bound) and the registry's own record says not published. In the builder. P4, P7.
- **H-A-270 (a pipeline's passed end is taken again when a mover's version arrives, 15:31)** Met live: the seat
mesh-wants' holder skips a passed pipeline's mover whose version is not yet written (the runner writes it at the
release's version-released, which waits on H-A-269), silently, so a passed pipeline wrote no want. The holder says
the skip in the journal, and takes the end again when the version arrives (on the version-released event it already
hears, or the runner says pipeline-ended again when a version is written after the end: the runner's side is the
rule, one event); the assigner keeps one handler for registration-changed. In the runner and the holder. P4,
P-NATS-native.