docs(design): membership & identity convergence — nessie rides agent-mesh; routing corrected

[hartsock]

Summary

Resolves the "coordinate swarm-join with agent-mesh before building a router" decision — grounded against the real, present-day types of agent-mesh and agent-bridle (read directly), plus the live newt-web/newt-agent membership story (web-authz epic newt-agent#1354 / #1365, agent-bridle 0.7.13).

New docs/design/membership-and-identity.md pins a three-layer stack in which nessie invents no identity and no membership:

Layer	Owner	Type
L1 operator	agent-bridle	HumanPrincipal { issuer, subject, email, groups }
L1↔L2 edge	agent-bridle	PrincipalBinding (signed, mesh AgentKey, Fingerprint, issued_generation)
L2 peer + membership	agent-mesh	AgentKey/Fingerprint/CertChain/MeshSigner; auto-team under one user_fp
L3 content-location	nessie	ContentRouter — the one net-new primitive

nessie's opaque identity newtypes map straight onto mesh types: SignerId ← Fingerprint bytes; SignatureVerifier/AttestationSigner ← AgentKey/MeshSigner (retiring DevSelfSigner/TestKeyring from non-test paths); k-of-n = k distinct Fingerprints under one user_fp; attestation GC by issued_generation + CertChain::verify_at(gen) — which answers the design's open "attestation garbage collection" question. Swarm-join = mesh-join (ensure_trustable handshake); no roster.

The correction

The fleet's fabric is iroh-QUIC + mDNS + cert-chain membership — NOT NATS, NOT libp2p (async-nats is only a comparison dev-dep in agent-mesh). So the earlier "settled: hybrid NATS + Kademlia, both first-class" is revised:

• Primary = a mesh-native Bus router (signed provider records on Topic(user_fp, "nessie/providers/v1")).
• NATS = a legitimate alternative for NATS-running environments, behind the same seam — not the default fabric.
• Kademlia/DHT = later, nessie-specific open-swarm variant (the mesh itself declined a DHT).

The ContentRouter seam is unchanged; only the ranking of implementations moves toward the fabric that exists. p2p-cas-swarm.md's routing section, tracker rows, closing prose, and the attestation-GC open question are all updated to match.

Convergence bonus flagged: nessie's AC attestation-CRDT is structurally the same primitive as the unbuilt agent-mesh-store RFC (per-writer signed BLAKE3-chained logs + head-vector gossip) — an opportunity to share one signed-log-over-a-bus-topic substrate rather than build two.

Test plan

Docs-only (no code). just check passes trivially (no Rust changed); the design tracker + prose are internally consistent and cross-linked.

Note for review

This revises a previously-"settled" routing decision based on grounding — surfaced explicitly for your call. No routers are built here; this is the coordination the "agent-mesh first" choice asked for. Implementation waits on a running mesh + the shared-log conversation with agent-mesh.

risk:low — documentation only.