route local LLM to qwen3.8 on .128 (was gemma4:26b); think:false + num_ctx 8192

src/server/llmClient.ts

@@ -1,7 +1,7 @@
 // Shared local-LLM client — Ollama on shadow-death (.128).
 // Single source of truth for host/model across the whole app.
 export const OLLAMA_HOST = process.env.OLLAMA_HOST || 'http://10.30.20.128:11434';
-export const OLLAMA_MODEL = process.env.OLLAMA_MODEL || 'gemma4:26b';
+export const OLLAMA_MODEL = process.env.OLLAMA_MODEL || 'qwen3.8:latest';
 
 export interface LocalModelOptions {
   formatJson?: boolean;
@@ -18,6 +18,8 @@ export async function callLocalModel(prompt: string, opts: LocalModelOptions = {
     options: {
       num_predict: opts.numPredict ?? 4096,
       temperature: opts.temperature ?? 0.2,
+      think: false,
+      num_ctx: 8192,
     },
   };
   if (opts.formatJson) body.format = 'json';