fix: UTF-8 terminal rendering — atob() decodes as Latin-1, not UTF-8

atob() returns a "binary string" where each byte is a Latin-1 char code.
Multi-byte UTF-8 sequences (box-drawing, em dashes, arrows) were split
into separate Latin-1 codepoints, producing mojibake. Now reconstructs
the raw byte array and decodes via TextDecoder('utf-8') before writing
to xterm.js.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

frontend/src/composables/useTerminal.ts

@@ -114,8 +114,14 @@ export function useTerminal(sessionId: string): UseTerminalReturn {
       }
 
       try {
-        const decoded = atob(b64data);
-        terminal.write(decoded);
+        // atob() returns Latin-1 — each byte becomes a char code 0x00–0xFF.
+        // We must reconstruct the raw bytes, then decode as UTF-8.
+        const binaryStr = atob(b64data);
+        const bytes = new Uint8Array(binaryStr.length);
+        for (let i = 0; i < binaryStr.length; i++) {
+          bytes[i] = binaryStr.charCodeAt(i);
+        }
+        terminal.write(new TextDecoder().decode(bytes));
       } catch {
         // Fallback: write raw if not valid base64
         terminal.write(b64data);