fix: A2A interop with real external agents (CLEAN-97)

[maksymhryzodub-prog]

Jira: CLEAN-97

External A2A agents mostly did not work with Ranch. The import and delegation circuit was fine; the interop was not. This fixes the four critical and three minor defects found by running real public agents through the real code, plus two more honesty gaps the same run surfaced.

Before and after

The same 41 public A2A 1.0 agents from a2aregistry.org, no auth, through the real PeerService.previewByUrl and DelegationService.run with production SSRF guards on:

	Before	After
Imported	40	40
Answered with content	6	36
Answered, but empty	3	0
Rejected: "answered without a task"	23	0
Skipped: first interface not JSON-RPC	4	0
Failed with a misleading cause	4	0
Failed, peer's own reason passed on	0	4

The 4 remaining failures are agents that demand structured input instead of text, and they now say exactly that. The five spellings of our own agent's URL that previously slipped past the self-import check on production are all refused now, before any request.

What changed

Critical

• Message replies accepted. A blocking SendMessage may return a Message instead of a Task; most public agents do. The client returns either and delegation reads both.
• The card's interface URL is vetted at import and external refresh. SSRF-guarded and refused when it points back into this installation. Existing rows fail delegation with PEER_ADDRESS_REFUSED instead of "it answered with an error".
• Self-import check compares a normalized address, not a string prefix: host case, trailing dot, port, decoded and collapsed path, scheme ignored. A card whose interface names our base is refused whatever host served it.
• Interface choice: the first JSON-RPC interface on 1.0, the same rule in import, delegation and the console's address line. A 1.0 card without JSON-RPC is refused with PEER_BINDING; an existing one fails with PEER_UNSUPPORTED.

Minor

• Reply text from every part. Data parts become compact JSON, links stay links, binary is named but not inlined; empty artifacts fall back to the task's status message. Only when nothing is readable is the reply recorded as an explicit empty answer, and the tool tells the model not to invent one.
• A 0.3 card reports its version instead of "not an agent card".
• Card URLs that already name a .json document are kept as is, so /.well-known/agent.json and custom paths are no longer mangled.

Found during verification, same class

• A JSON-RPC error sent with a 4xx status now passes the peer's reason on instead of "could not be reached".
• Redirects are still never followed, but redirect: 'manual' reports them with their target, so the operator knows which address to paste instead of seeing "fetch failed".

Verification

Check	Result
API tests	1073 passing, 83 suites
Peer slice tests	274 passing, including new a2a.types.spec.ts
API tsc --noEmit	clean
Admin and app nuxt typecheck against a freshly generated SDK	clean
ESLint on changed API files	clean
Live run against 41 public agents	table above

Notes for review

• Generated SDKs are not committed. Only an API DTO description changed. Regenerating on this machine reformats every generated file, a 17k-line formatting diff, and the consoles regenerate the SDK on prebuild and pretypecheck anyway.
• Pre-existing, not from this branch: npx nuxt typecheck in admin against the committed SDK fails in the knowledge slice, because CLEAN-96 added the retrying status without regenerating the committed client. bun run typecheck regenerates first and passes.
• Not verified on production: the manual flow needs this deployed. The import, preview and delegation paths above ran with the real code against real agents; the chat side and the armed indicator are unchanged.
• Version bumped to 0.3.54.

🤖 Generated with Claude Code