fix(web): propagate mid-stream API errors and raise max_tokens

- Streaming loop swallowed errors: a mid-stream error event (e.g.
  overloaded_error) was thrown inside the same try/catch used to skip
  unparseable SSE lines, so it was silently ignored and the run reported
  "Done." with truncated output. Separate JSON parsing from event handling
  so real errors surface to the user.
- Raise max_tokens 4096 -> 8192 to avoid truncating long skill outputs.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

web/app.js

@@ -170,7 +170,7 @@ async function run() {
       },
       body: JSON.stringify({
         model: el('model').value,
-        max_tokens: 4096,
+        max_tokens: 8192,
         stream: true,
         system,
         messages: [{ role: 'user', content: userMessage }],
@@ -193,15 +193,18 @@ async function run() {
         if (!line.startsWith('data:')) continue;
         const payload = line.slice(5).trim();
         if (!payload || payload === '[DONE]') continue;
+        let evt;
         try {
-          const evt = JSON.parse(payload);
+          evt = JSON.parse(payload);
+        } catch (_) {
+          continue; // skip an unparseable / partial SSE line
+        }
         if (evt.type === 'content_block_delta' && evt.delta && evt.delta.text) {
           acc += evt.delta.text;
           renderMarkdown(out, acc, true);
         } else if (evt.type === 'error') {
-            throw new Error(evt.error ? evt.error.message : 'stream error');
+          throw new Error(evt.error ? evt.error.message : 'Stream error from the API.');
         }
-        } catch (_) { /* ignore partial */ }
       }
     }
     renderMarkdown(out, acc, false);