Authentication for ACP over Streamable HTTP

[mgoldsborough]

The Streamable HTTP transport makes remote agents practical, and the RFD leaves auth "layered on top". That works for one deployment, but every agent ends up picking its own scheme. Goose's serve uses a fixed shared secret in X-Secret-Key or ?token=, and a host that dials several agents has to carry each one's scheme.

We dial agents over HTTP from a host that already signs short-lived tokens for its other services. Each connection carries a bearer JWT that the host signs, with an audience for the agent pool, checked against a published key set. Keys rotate by kid and nothing secret is shared with the agent. It's the resource-server pattern MCP's authorization spec already uses.

Could the transport say how an agent declares what it accepts? For example, a bearer token in Authorization, plus where to find the issuer's keys, so a host knows what to send without reading each agent's docs. This sits next to the connection-scope point on #1977: if model credentials become per connection, the connection itself has to be authenticated first, and both should use the same mechanism.