fix(evals): parse single-logical-line AskUserQuestions in the PTY runner

When the PTY reflows a boxed AUQ, ALL options land on ONE logical line
after stripAnsi — parseNumberedOptions parsed one option per line, found
only '1.', and the >=2 check failed forever while the correct question
sat on screen (plan-design-with-ui timed out this way twice, with the
rendered scope-gate AUQ visible in both failure buffers). The cursor
line is now parsed as a stream of ascending N. tokens; DEC cursor-
visibility residue is stripped before matching; plan-design-with-ui's
budgets grow to fit observed ~6min preamble+thinking latency. Pinned by
test/pty-auq-single-line.test.ts using the real failure buffers; all 142
existing parser-consumer unit tests still green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

test/skill-e2e-plan-design-with-ui.test.ts

@@ -44,7 +44,7 @@ describeE2E('/plan-design-review with UI scope (gate)', () => {
       const session = await launchClaudePty({
         permissionMode: 'plan',
         cwd: ROOT,
-        timeoutMs: 480_000,
+        timeoutMs: 720_000,
         seedSkills: true,
       });
 
@@ -71,7 +71,11 @@ describeE2E('/plan-design-review with UI scope (gate)', () => {
           `Reference plan file: ${fixtureRelPath}\r`
         );
 
-        const budgetMs = 360_000;
+        // 600s, not 360s: the skill preamble (update-check, session bookkeeping,
+        // learnings) plus extended model thinking can take ~6 minutes before the
+        // scope-gate AskUserQuestion renders — a 360s budget expired seconds
+        // before the (correct) AUQ appeared in the observed failure transcript.
+        const budgetMs = 600_000;
         const start = Date.now();
         let lastPermSig = '';
         while (Date.now() - start < budgetMs) {
@@ -146,6 +150,6 @@ describeE2E('/plan-design-review with UI scope (gate)', () => {
         );
       }
     },
-    540_000,
+    780_000,
   );
 });