refactor(rustd): a connector is a descriptor, and a fault is not a posture

Two names Indy caught, and one of them was a regression against the
Zig's own rule rather than a port of it.

`credentials/integration.zig`'s header states RULE CFG outright — "a
connector is a descriptor in REGISTRY, not a branch in the broker's
mint dispatch" — and `mintsOnDemand(registry, id)` takes its registry as
a parameter so a test can inject a fake. The enum collapsed both away.
It read well at five connectors and cost three edits at six: the
variant, the spelling, and the supply, each in a different function,
each easy to get right and easy to forget.

Now it is `Connector` + `Connectors` + `Descriptor` + `Registry`, which
is `afd_fleet_runtime::provider`'s shape byte for byte. That module
solves the same problem for webhook signature schemes — declared data
now, a crate-backed client later, one trait so callers cannot tell the
difference — and two registries in one workspace answering the same kind
of question should not be two different designs.

Adding a connector is one entry in `DECLARED`. The mint exchange
attaches later as a second trait a descriptor implements; nothing in
this file moves when it does, and no caller learns about it. That is
the reason the seam is a trait now rather than after the broker lands.

`Supply` is two arms rather than a bool, because the two answers are not
"yes" and "not yes" — they are two delivery mechanisms, and a reader
meeting `supply()` should not have to remember which way round a bool
went. The exhaustiveness the enum gave up is bought back by tests: every
declared name resolves, no name is declared twice, and only `static` is
inline.

`Mintable.integration` is the connector's NAME, not a `&dyn Connector`.
That value travels on the execution policy and the broker resolves it
again through its own registry; borrowing would tie a whole lease to the
registry's lifetime for a string it is about to serialize.

Separately: `admit/posture.rs` is now `admit/fault.rs`. `Posture` is
already taken by the money path — it is `billing.usage_ledger.posture`'s
own spelling, so it keeps the name — and the module beside it meant
`admit/mod.rs` imported `posture::PAYER` and `Posture` on adjacent lines
for two unrelated questions: who pays, and what a dead datastore means.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>