The router reports a shed request as unavailable after a full fleet #2127

Follow-up to #1896. Parent: #1664.

Context

On the router's atenet.router.route.duration histogram, the ate.router.outcome (ateattr.RouterOutcomeKey) attribute is computed by classifyOutcome(err) in cmd/atenet/internal/router/extproc/metrics.go.

Prior to PR #1929, classifyOutcome(err) mapped every envoy_type.StatusCode_ServiceUnavailable (HTTP 503) *extproc.ReqError without a wrapped gRPC status code to no_capacity (ateattr.RouterOutcomeNoCapacity). Because both the ingress and egress handlers record atenet.router.route.duration, non-capacity 503s (such as egress denials or policy/credential lookup failures) were falsely reported as no_capacity.

PR #1929 fixed that over-reporting by updating classifyOutcome so that no_capacity is reported only when status.Code(err) == codes.ResourceExhausted (unwrapping through ReqError.Cause), while bare 503 ReqErrors without a ResourceExhausted cause fall through to unavailable (ateattr.RouterOutcomeUnavailable).

Expected Behavior

A request that the router sheds while the resume of its actor is retrying because of ResourceExhausted reports no_capacity. A shed request with a different cause continues to report unavailable. The unavailable brief in docs/metrics/registry/metrics.yaml and the comment on the router outcome values in internal/ateattr/ateattr.go describe the new behavior.

Discussion: #1929 (comment)