T014 single tailnet DNS endpoint — live verification
Publication redaction: Host-specific Tailnet IPv4, Tailnet IPv6, and MagicDNS identities were replaced with the repository's synthetic IPv4, explicit IPv6 placeholder, and example DNS name; HTTP results and event ordering are unchanged.
Captured: 2026-07-13T20:16:41Z  host: primary-node (RTX 5090)
MagicDNS (tailscale status --json .Self.DNSName): node-a.example.ts.net.
Tailnet IPs: 100.64.0.10 / <tailnet-ipv6>
All requests below target http://node-a.example.ts.net:8000 (the MagicDNS name), never a raw IP.

== GET /v1/models (authed) ==
HTTP 200
== GET /v1/models (no token) ==
HTTP 401

== GET /healthz (unified route surface) ==
{"status": "ok", "dialects": ["anthropic", "openai"], "routes": ["/v1/chat/completions", "/v1/decisions", "/v1/embeddings", "/v1/messages", "/v1/rerank", "/v1/route"]}

== POST /v1/embeddings (qwen3-embedding-0.6b) ==
HTTP 200
object: list model: qwen3-embedding-0.6b dim: 1024
embedding[:3]: [0.0365, 0.0477, -0.0116]
usage: {'prompt_tokens': 6, 'total_tokens': 6, 'completion_tokens': 0, 'prompt_tokens_details': None}

== POST /v1/embeddings (unknown model -> clean 404, names served model) ==
{"error": {"type": "model_not_found", "message": "unknown embedding model '__unknown__'; this router serves embedding models: 'qwen3-embedding-0.6b'. embedding requests are routed by model name only and never fall through to chat routing."}}