Infra decoded real tokens from both clients, and two of the three facts changed
this code.
A machine token carries NO sub at all: it names its caller by client_id, and the
identity is that client. A person's token carries sub, the provider's stable
identifier for the person. The subject is now derived explicitly -- sub when
present, else client_id -- and a verified token asserting neither is refused
rather than accepted as an anonymous caller, because accepting it would let any
token the provider ever signed act as whoever the binding happened to hold. A
machine identity is therefore per client registration, one identity per client,
which is what makes per-agent identity a configuration decision rather than a code
change.
Scopes are read from either claim shape. The standard is a space-delimited scope
string and this provider also emits an scp array; reading only one shape yields no
scopes at all, which is worse than reading none deliberately.
The audience-at-authorisation-endpoint trap is pinned by a test: a token minted
without the requested audience arrives with aud empty, and the audience check
refuses it. That refusal is correct and the fix is the request the human flow
builds, not a weaker check here -- a test now fails if anyone weakens it.
The PRD records both as constraints: a person's client and an agent's client are
registered separately because a shared registration collapses the distinction the
attribution rests on, and a person's token requests the audience at the
authorisation endpoint as well as the token endpoint.
Verified: task check exit 0; internal/infrastructure/oidc passes, including
subject derivation for both token shapes and every refusal.