Design 47 H2: H-A-430; words of H-A-429

This commit is contained in:
2026-11-22 01:25:27 +01:00
parent 3b29f26c77
commit c1a5b310fe
@@ -484,3 +484,31 @@ The lead's one-line answers to the questions H2's develop and review raise (impl
workspace tree's root lays the modules where the Go build module's tests expect them (BM2's layInstall). The install
writes the package.json npm ci needs from the lock's root entry (name, version, the four dependency maps, no scripts)
and refuses a lock with no root entry (lockfileVersion 1), naming it. P3, P-failure-loud.
- **H-A-428 (the validator's catalog-wide tests read the group's catalog head, 01:24)** §814, I6. A group's check
judges every member against the group's own heads (ADR 0573 §2), so mesh-judge's tests over the whole catalog read it
through internal/beside from the group member's head when the catalog is in the group (the sibling placement the
check gives, MESH_CHECK_BESIDE, as mesh-catalog's own checks read mesh-sdk and mesh-delivery beside them), and from the
catalog's main only when it is not; the captured copy testdata/beside is refreshed from M1's head (feat/47-H2-c,
merged into mesh-catalog feat/47-H2) in the same commit as the two rules. The cycle (a validator refusing `class` holds M1
off main, a check failing without M1 holds the validator off) is the group's to break, never a rule judged "without the H2
rules". design-it-as-it-should-be, ADR 0573 §2.
- **H-A-429 (a want's run is compared within its branch; the private network manifest's class, 01:27)** A4, I6. (1) A want's
`run` is read with the branch the want names: the engine refuses a lower run of the same branch only; a want of
another branch is a placement the assigner made on purpose and replaces whatever stands. A want with no pipeline
carries the newer of the kept trunk source and the held want's run (the developer's choice, confirmed). (2) The
private network's manifest the mesh provides itself (mesh-sdk's `overlay.Manifest`, mesh-wireguard) claims a seat, so
it states `class: core` (H-A-394's rule); a one-line mesh-sdk commit by the indexer's developer, who meets it, citing
this answer; the captured catalog files of the validator's tests are refreshed from M1's head (H-A-428). P3.
- **H-A-430 (C4's generation and first mint; the runner's step words; I6's order; finding-events on every module,
01:28)** (1) C4: `<generation>` is the vault's generation of module/gitea/<account> that the own-put will make (get's
generation plus one; 1 when the vault holds none); no store of its own. The first mint deletes by id the account's
tokens under the old fixed names (mesh-tools, mesh-packages) as the tokens it replaces; a later mint finds the
replaced token by its forge name `<account>-<generation>`. (2) runner_movers' five words: could-not-run and
did-not-build read failed; declined, superseded and cancelled read not-run (the developer's mapping, confirmed). (3)
I6 is committed when the catalog's feat/47-H2 (with -c merged) carries every manifest the rules judge: the build
modules' and distribution's class and caps (BM/RG rows), and `finding-events` at max-mib 8 on every module that binds
one, not only the delivery modules (H-A-411 (1) read so: eleven more rows in M1, the catalog-c developer); the
captured copy is refreshed from that head then, and I6 stays on its side branch until. (4) Two refusals the H2 rules
did not make are catalog faults fixed in M1: nftables' setting `allowances-max` is marked "runtime": true, and
mesh-recovery's claimed seat is one the group's seat list holds (its definition read from the group's heads). P3, a
rule states how it is checked.