Design 47 G: the lead's answers A-454a, A-456..A-462 (the delivery path after a passed pipeline) #643

Merged
mesh-tools merged 1 commits from design/47-G-lead-answers-57 into main 2026-11-17 00:05:32 +00:00
@@ -3,7 +3,7 @@ layer: to-be
status: in-progress
design: 47
milestone: G
updated: 2026-11-13
updated: 2026-11-17
decisions:
- "ADR 0667"
---
@@ -1343,3 +1343,55 @@ the principle that settles it, cited as "design 47 G A-n".
publish as §588's publish-again does. (Live, 2026-11-17 00:49: the tool runner's (`node-tools`) 0.1.13 stayed reserved for build b5c7a048,
whose entry was gone, so the passed build 7df8e170 of main.984 reserved nothing, was never released, and its wants
carry no parts; a done key cannot be built again.)
- **A-454a** (A-454; principle 3) The A-454 exception applies only to the direction "fetched and not in the want's parts":
a digest the want's parts name always counts as fetched when the composition fetches it, whatever marks its
resource carries; a fetched digest the parts do not name is counted only when it is the mesh's own (named through
the artifact store or the registry seat, and not an upstream copy). (From the review of A-454: a mesh-built image
whose seat holder is not on record carries no `image-seat`, so skipping it would hold the module on its own part.)
- **A-456** (§369, §588, A-455; principle 3) A pipeline's build step passes only when the build it judges has a
reserved version for its fingerprint. A build that ended done with no reservation fails the step loudly ("built at
<fp>, but no version is reserved for it"), naming `builder_reserve_done` as the recovery, so no pipeline says
"passed" for a build that can never be released. (Survey 2026-11-17 00:58: the tool runner's build 7df8e170 passed
main.984 and was never released.)
- **A-457** (§588; principles 1, 3 and 4) A version's publish is work the builder owns until `version-released` is
written for it: every reservation of a done build is a publish item on the builder's own work queue, asked when the
build ends, redelivered after a failed or unanswered publish with a growing pause (a setting, principle 11), and
acknowledged only once the release is written. At its start the builder puts on the queue every reservation whose
build is done and that has no release, so what stood before this change moves too. A publish that fails its
setting's number of tries is said loudly (log and condition) and stays readable in the builder's tools; it is never
waited on a `running` or `manifest-reads-changed` event any more. (Survey 2026-11-17: 115 reservations, 108 built,
3 released.)
- **A-458** (§354's composition step; principle 3) The composition step judges by what the composer said: when it judged
no candidate (every verdict "not judged"), the step ends "could not run" with the composer's words, never "passed";
a step passes only when at least the candidates it was asked are judged and none refused. The asker (the assigner)
names the seat list in the request's `seats` from the seat list it holds (ADR 0636), so the composer has what it needs
to judge. (Survey 2026-11-17: 86 of 141 pipelines "composition could not run", and the code read "not judged" as a
pass.)
- **A-459** (the assigner's rule 7; principles 10 and 3) Rule 7's "composed" is asked per module: a module waits only
on the composition of its own have at the fingerprint it waits for, on that node, never on whether the node composed
anything else. One module that does not compose on a node never holds another module there. (Survey 2026-11-17:
`composed(node, module, fp)` read the node's own have and ignored module and fingerprint, so one node that composes
nothing held every module on it.)
- **A-460** (A-452, ADR 0539; principles 10, 3 and 2) A node-engine's install of a build of itself that fails (the
version's folder cannot be made, its archive not fetched) costs the node-engine's own have only: it stays on its
current build, its have says why with the error's words, and every other module of the node composes. The
node-engine makes its versions folder where its own unit lets it write; a refusal there ("operation not permitted")
is a fault of the node-engine's install and is fixed in it, not on the machine. (Live, ace 2026-11-16 23:33: the
node composed nothing after `mkdir .../versions/88216c0e9496.unpacking: operation not permitted`; its want for the
node-engine names an older build, d9e5d175, because no pipeline of the node-engine passed at a newer one; that want
moves with this change's own passed pipeline, by §29.)
- **A-461** (A-452, A-407; principles 2 and 10) Where a module on this node needs another node's module (a store
address, a provider's endpoint) and that module runs a build that has no entry, the node lays it from the entry of
the build its want names, which is the mesh's truth of what it runs there. Where neither build has an entry, the
holder stays name-only and only the modules that need its address wait, as A-452's consumers do, with the same
words; the node is never refused whole for it. (Live, ace 2026-11-17: ace laid the artifact store's holder
`distribution` name-only, novox running 44a32f00 with no entry while its want names 4133e115, which has one, and
refused its whole composition, "the artifact store has no address".)
- **A-462** (§29, A-428; principles 2, 5 and 3) A want of a module that names no version and no parts while the build
it names has a released version is completed by the assigner in one pass of its own: the want keeps its
fingerprint (nothing moves, so §29's passed pipeline is not needed for it) and gets that version and its parts. The
pass runs at the assigner's start and on every `version-released`, and says per want what it filled. A want whose
build has no entry is not filled: the indexer is asked to take that build in again (`indexer_take_in_pending`), and
the want says it waits for the entry. (Survey 2026-11-17: 40 to 80 modules per machine on wants from before G with
no version or parts, `entry-missing` or "want names no digest", and a want was rewritten only after a passed
pipeline, which waited on A-458.)