Network-AI: Empty default secret still authorizes all requests (Incomplete fix for CVE-2026-46701)

An attacker who can reach the MCP SSE server port (e.g. via an SSRF on the host, a non-loopback bind in Docker/remote deployments, or a direct curl request) can invoke all 22 MCP tools — including `config_set`, `agent_spawn`, `blackboard_write`, and `token_*` — without any credentials [ref_id=2]. The server defaults to an empty secret, so `_isAuthorized()` unconditionally returns `true` and no `Authorization` header is required [ref_id=1]. The CORS wildcard was fixed in 5.4.5 to only allow localhost origins, eliminating the browser drive-by vector, but non-browser callers remain fully unauthenticated [ref_id=2].

The flaw resides in `bin/mcp-server.ts` (line 89) where the secret defaults to an empty string (`process.env['NETWORK_AI_MCP_SECRET'] ?? ''`), and in `lib/mcp-transport-sse.ts` (line 254) where `_isAuthorized()` returns `true` when the secret is falsy. The `listen()` method only prints a warning on non-loopback binds instead of refusing to start [ref_id=2].

The fix in v5.7.2 makes two changes: `_isAuthorized()` now returns `false` (fail closed) when the secret is empty, and `listen()` rejects with a hard `Error` if `McpSseServerOptions.secret` is empty, preventing accidental open deployments [ref_id=3]. This ensures that an empty secret is treated as "deny all" rather than "allow all", closing the authentication bypass that survived the earlier CORS-only fix.