perf(daemon-client): persistent remote clients re-run the /health probe on every command #2650

Problem

ensureRemoteDaemon (src/daemon-client/daemon-client-lifecycle.ts:174) calls readRemoteDaemonHealth on every sendToDaemon, so one proxied command is two round trips:

GET  /health
POST /rpc

For a fresh-process CLI that is one probe per process and unavoidable. For a persistent client (the MCP server, the benchmark's persistent-client mode) it is a full extra round trip per command. It matches the ~2.5 round trips per snapshot the #2198 slice-B benchmark measured (persistent-client medians 36 / 93 / 233 ms at 0 / 20 / 80 ms RTT).

Found in #2381 while asserting #2198's round-trip line in #2387; it predates the bridge and is unrelated to it.

Required behavior

A persistent client pays the ADR 0006 health probe once per daemon identity, not once per command. A fresh process still probes before its first command. Re-probe on a transport failure or a changed daemon identity (base URL, token, pid), not on a timer alone.

Completion

• Persistent-client benchmark shows one wire request per command after the first.
• A daemon restart between two commands is still detected (compat and version-skew checks in daemon-client-health-compat.test.ts keep passing).
• Fresh-process CLI behavior unchanged.