v1.6.0 - dual-node runtimes, usage stats, recipe search

Folds in 1.5.0-1.5.2, which were bumped in pyproject but never tagged or
released, so this covers everything since v1.4.0.

Features
- vllm-dual runtime: TP=2 serving across two Sparks over CX7
- vllm-dual-ray runtime: Ray-based cross-node TP (MiMo-V2.5)
- llama-dual runtime: GGUF split across both Sparks over llama.cpp RPC
- per-recipe shm_size and ulimit_memlock for vLLM containers
- usage statistics tracking, then token tracking + visual dashboard
- recipe view filters, keypress TUI, and search
- vLLM mirrors container logs to running/<name>.log like gguf; container ID
  written to ai-models/running for unified PID logging

Fixes
- cmd_serve never detected a crashed llama-server, defeating
  restart=on-failure (model silently unreachable under systemd)
- cluster-aware config sync for opencode/hermes/grok
- deduplicate cluster models; fix remote model filtering
- shared dual port (8888) made every dual recipe report as running
- prefer SSH-alias-resolved host over the peer's flaky mDNS self-report
- search TUI key, dual TYPE display, host misattribution, dual mount
  asymmetry

Other
- ai-models is no longer tracked as a git submodule
- record_stop uses an in-memory cache: O(1) lookup, zero I/O

v1.4.0 — vLLM entrypoint: and env: keys

lmswitch 1.2.0

v1.1.0 — multi-target config sync

Sync currently-serving models to opencode, hermes, and grok (selectable at
init). Hermes gets a sticky active model, auxiliary.vision wiring, and a
custom_providers entry per model (discover_models: false) so they're all
selectable from /model without the picker hanging on live probes.

lmswitch v1.0.0

Local LLM switcher for GGUF (llama-server) and vLLM (Docker), driven by
per-model YAML recipes. First stable release.