infra: migrate GPU workstation to Ubuntu, deploy vLLM-Hermes + voice stack

- Migrated primary workstation (192.168.86.27) from Windows/WSL2 to Ubuntu 24.04 (2026-05-28)
- Upgraded Ollama 0.15.6 → 0.24.0 (all 21 models preserved)
- Fixed Watchtower crash loop (DOCKER_API_VERSION=1.44)
- Fixed Open WebUI OLLAMA_BASE_URL to use Docker service name
- Deployed Speaches :8000 (CUDA, STT/TTS for VoiceCoach)
- Deployed LiveKit :7880/:7881 (WebRTC for VoiceCoach)
- Deployed vLLM-Hermes :8001 (Hermes 4.3 36B AWQ, tool-call-parser hermes)
  - Endpoint: http://192.168.86.27:8001/v1, model: hermes
  - Tool-call test passed: structured tool_calls array confirmed
- Added vllm-local provider to voicecoach-openclaw and margaret OpenClaw configs
- Added OPENAI_API_BASE_URL to Open WebUI for Hermes access in chat UI
- Updated CLAUDE.md: removed WSL2 references, added GPU workstation section

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

CLAUDE.md

@@ -36,7 +36,7 @@ hf-pdocker-01 (23GB RAM, i7-7700T, NVMe, HD 630 GPU) -- RESTORED (2026-03-08, ha
 ├── [COMPOSE] Farm Manager API (:8888) -- dashboard + API server
 ├── [COMPOSE] Farm Agent (:8889)
 ├── [COMPOSE] Rongorongo decipherment engine -- 2 containers (rongorongo-web :8181, rongorongo-api). Server-synced state at /mnt/docker-data/configs/rongorongo/data. Traefik basicAuth on rongorongo.hobbs.farm.
-├── [REMOVED] VoiceCoach -- moved to GPU machine (192.168.86.150) 2026-04-04
+├── [REMOVED] VoiceCoach -- moved to GPU machine (192.168.86.27) 2026-04-04, now Ubuntu
 ├── [COMPOSE] Node-exporter (:9100) -- standalone with pid:host
 ├── [COMPOSE] Telegraf -- metrics collection to InfluxDB
 ├── [CRON] sync-frigate-to-nas (daily 2am), cleanup-frigate-media (weekly Sun 3am), restart-frigate (daily 4am)
@@ -51,6 +51,21 @@ bart (15GB RAM, i5-4570T, HDD)
 ├── [COMPOSE] Telegraf -- metrics collection to InfluxDB
 ├── [COMPOSE] Watchtower
 
+GPU workstation / Jeremy's PC (192.168.86.27, Ubuntu 24.04 — migrated from Windows 2026-05-28)
+├── RTX 3090 (24GB VRAM), i7-13700KF, 32GB RAM
+├── [COMPOSE] Ollama (:11434) -- 21 local models, v0.24.0
+├── [COMPOSE] Open WebUI (:8080) -- AI chat, points to local Ollama
+├── [COMPOSE] vLLM-Hermes (:8001) -- Hermes 4.3 36B AWQ, OpenAI-compat, tool-call-parser hermes
+│   └── Endpoint: http://192.168.86.27:8001/v1  model name: hermes
+├── [COMPOSE] Speaches (:8000) -- STT/TTS via faster-whisper + Kokoro, CUDA
+├── [COMPOSE] LiveKit (:7880/:7881) -- WebRTC for VoiceCoach, key: devkey
+├── [COMPOSE] Portainer (:9000, :8002)
+├── [COMPOSE] Prometheus (:9090), DCGM-exporter (:9400), node-exporter (:9100)
+├── [COMPOSE] Watchtower (label-enable mode)
+├── Compose file: /home/glyptic/Docker/compose/docker-compose.yml
+├── vLLM compose: /home/glyptic/Docker/vllm-hermes/docker-compose.yml
+└── Model cache: /home/glyptic/Docker/data/models/
+
 hf-pdocker-02 (7.6GB RAM, i7-6700T, HDD)
 ├── [COMPOSE] Unifi + unifi-db (MongoDB)
 ├── [COMPOSE] Gitea + gitea-db (PostgreSQL)
@@ -69,9 +84,11 @@ hf-pdocker-02 (7.6GB RAM, i7-6700T, HDD)
 
 ### Environment Context
 
-You are working in a **WSL 2 environment** on Windows. Docker is NOT directly available in WSL - all Docker operations must be performed by connecting to the remote nodes via SSH.
+You are working on a **native Ubuntu 24.04 workstation** (previously WSL2 on Windows — migrated 2026-05-28). This machine IS the GPU inference box (RTX 3090, 192.168.86.27). Docker is available locally and runs the inference stack. All homelab node operations (hf-pdocker-01, hf-pdocker-02, bart) still require SSH.
 
-**Critical:** Docker is not available locally in WSL. All Docker operations must be performed by connecting to the remote nodes via SSH.
+**Local Docker on this machine** runs: Ollama, Open WebUI, vLLM-Hermes, Speaches, LiveKit, Portainer, Prometheus, DCGM-exporter.
+
+**Critical:** For homelab nodes, use the `docker-swarm-ops` agent via SSH. For local inference services, Docker commands run directly.
 
 ### Storage Architecture