Answer A-490a and A-490b: where the package holders' publish cases live; a rebuild reserves its version afresh

Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
This commit is contained in:
2026-11-19 17:02:43 +01:00
parent 9bde22296e
commit 1dd4daf184
@@ -2055,3 +2055,15 @@ the principle that settles it, cited as "design 47 G A-n".
a package is kept anywhere else. A package whose build's publish was refused is built again
(`builder_build_again`), never published from a copy. (Live 2026-11-19: the sdk's build publish refused at 12:25 by
the retired `callers` read; the builder's no-parts republish refused for the parts form.)
- **A-490a** (A-490, §680; principle 9) The cases every package holder is held to for the no-parts publish (published
when the forge holds that digest at the version, refused `not stored` naming the digest otherwise, also when the forge
holds the version under another digest) live in mesh-delivery-sdk beside the build modules' conformance kit, as §680
places that kit: they are the contract of the seats `go-package-registry` and `npm-package-registry`, which more than
one module fills, so each holder's own tests run them, every seat it fills. (From review round 1 of the gitea
no-parts publish fix.)
- **A-490b** (A-490, §369, §412; principle 2) A done trunk build whose reserved version is unreleased is built again
under a reservation made afresh, as a build takes over an older unreleased reservation (A-455): next patch above the
highest the catalog's `versions` answers, never the standing number kept, so a reservation made while the catalog's
first fill had not yet read the forge's tags (§412) does not outlive it. A trunk build whose reservation cannot be made
fails, and builds nothing. (Live 2026-11-19: the sdk reserved 0.1.0 while the forge holds `go/v0.1.16`, because the
catalog's `versions` answers no version for any module.)