feat: Spark Workbench everywhere, autonomous Hadoop offload & LLM masking-aware

- Data Hub with Hadoop tab (HDFS/Iceberg browser, Spark, pipeline)
- Databricks-style Lakehouse Workbench (Trino engine, live exec matrix,
  materialize to Iceberg/S3); reused & embedded in every source-DB UI
- HDFS -> Kafka -> Spark -> Iceberg/S3 pipeline; WebHDFS hostname resolver
- Data Flow master pulse switch (Run/Pause/Stop) gating animated edges
- Data Custodian autonomous Hadoop offload loop (batch counterpart to CDC),
  pulsing source -> HDFS edges; toggle in Data Flow
- LLM now autonomously aware of all latest platform changes (live platform
  context) and enforces masking policy: never reveals masked PII, still
  answers helpfully with aggregates/explanations

ui/src/context/ThemeContext.tsx

@@ -12,12 +12,12 @@ const STORAGE_KEY = 'atc-command-center-theme'
 
 function readStored(): Theme {
   const v = localStorage.getItem(STORAGE_KEY)
-  return v === 'dark' || v === 'light' ? v : 'light'
+  return v === 'dark' || v === 'light' ? v : 'dark'
 }
 
 export function ThemeProvider({ children }: { children: ReactNode }) {
   const [theme, setTheme] = useState<Theme>(() => {
-    if (typeof window === 'undefined') return 'light'
+    if (typeof window === 'undefined') return 'dark'
     return readStored()
   })