feat: localchat, a Go + Templ + HTMX chat app for local Gemma models

Streams replies over SSE from any OpenAI-compatible server (oMLX,
llama.cpp, Ollama, ...). Single binary that can install itself as an
OS service; docker compose bundles llama.cpp + Gemma 4 E2B.

compose.yaml

@@ -0,0 +1,44 @@
+# localchat + a local Gemma 4 E2B model served by llama.cpp.
+#
+#   docker compose up -d        # first start downloads the model (~3 GB)
+#   open http://localhost:3000
+#
+# Swap the model with LLM_HF_REPO, e.g. LLM_HF_REPO=unsloth/gemma-4-E4B-it-GGUF:Q4_K_M
+services:
+  llm:
+    image: ghcr.io/ggml-org/llama.cpp:server
+    command:
+      - -hf
+      - ${LLM_HF_REPO:-unsloth/gemma-4-E2B-it-GGUF:Q4_K_M}
+      - --no-mmproj      # text-only chat; skip the vision projector download
+      - --ctx-size
+      - "8192"
+      - --port
+      - "8080"
+    environment:
+      LLAMA_CACHE: /models
+    volumes:
+      - models:/models
+    healthcheck:
+      test: ["CMD", "curl", "-fs", "http://localhost:8080/health"]
+      interval: 10s
+      timeout: 5s
+      start_period: 15m   # first boot downloads the model
+    restart: unless-stopped
+
+  app:
+    build: .
+    image: localchat:latest
+    ports:
+      - "3000:3000"
+    environment:
+      LLM_BASE_URL: http://llm:8080/v1
+      LOCALCHAT_TITLE: ${LOCALCHAT_TITLE:-localchat}
+      SYSTEM_PROMPT: ${SYSTEM_PROMPT:-}
+    depends_on:
+      llm:
+        condition: service_healthy
+    restart: unless-stopped
+
+volumes:
+  models: