test(skillgym): require ref-driven SpringBoard interaction, not fixed labels

The case required the literal label="Add Widget" while the topic it
tests explicitly says SpringBoard labels vary and must come from a
fresh snapshot; it also didn't require the Calendar search action, so
a plan could pass while skipping the widget-gallery search entirely.
Give the contract explicit snapshot-derived refs (@e30/@e31/@e35) and
require pressing/filling those refs plus the Calendar search, and
forbid hardcoded English labels (Add Widget/Edit/Done) instead.

Verified locally against three synthetic finalOutput plans (good,
search-skipping, fixed-label) since the real runner needs a live
model.

Author: thymikee

test/skillgym/suites/agent-device-smoke-suite.ts

@@ -2596,22 +2596,31 @@ Treat the recovery message as a warning, not a fatal error. Use the exposed Sear
       'Agent Device Tester is already open on this session; its widget-ready state is prepared',
       'agent-device open com.apple.springboard --platform ios binds the session to SpringBoard for selector-driven commands; there is no separate widget/system command',
       'Entering SpringBoard edit mode needs one deliberate empty-space long-press on the home screen; there is no selector for it',
-      'The widget-gallery search-result row for "Calendar" currently returns an unlabeled accessibility node, so select it by coordinates read from a screenshot',
-      'SpringBoard control labels vary by iOS version and locale; discover them from a fresh snapshot -i instead of assuming fixed English strings',
+      'After edit mode, a fresh interactive snapshot exposes the Edit menu entry that opens the widget gallery as ref @e30',
+      'Inside the widget gallery, a fresh interactive snapshot exposes the search field as ref @e31',
+      'The widget-gallery search-result row for "Calendar" currently returns an unlabeled accessibility node, so select it by coordinates read from a screenshot instead',
+      'After choosing the result, a fresh interactive snapshot exposes the size-picker placement control as ref @e35',
+      'SpringBoard control labels vary by iOS version and locale; do not assume fixed English strings such as "Add Widget" or "Edit" — use the refs above, sourced from a fresh snapshot -i, instead',
     ],
-    task: 'Plan the commands to open SpringBoard, enter edit mode, search the widget gallery for "Calendar", add it using the documented coordinate fallback for the unlabeled search-result row, reopen Agent Device Tester to resume app automation, and close the session.',
+    task: 'Plan the commands to open SpringBoard, enter edit mode, open the widget gallery with @e30, search it for "Calendar" using @e31, select the result using the documented coordinate fallback for the unlabeled search-result row, place the widget with @e35, reopen Agent Device Tester to resume app automation, and close the session.',
     outputs: [
       plannedCommand('open com.apple.springboard'),
       /--platform ios/i,
       /snapshot\b[^\n]*-i\b/i,
       plannedCommand('longpress'),
+      /(?:^|\n)(?:agent-device\s+)?(?:press|click)\s+@e30\b/i,
+      /(?:^|\n)(?:agent-device\s+)?fill\s+@e31\b[^\n]*Calendar/i,
       /(?:^|\n)(?:agent-device\s+)?screenshot\b/i,
       RAW_COORDINATE_TARGET,
-      /label=["']Add Widget["']/i,
+      /(?:^|\n)(?:agent-device\s+)?(?:press|click)\s+@e35\b/i,
       plannedCommand('open com.callstack.agentdevicelab'),
       plannedCommand('close'),
     ],
-    forbiddenOutputs: [plannedCommandAlternatives(['widget', 'widget add']), /widget\s+add/i],
+    forbiddenOutputs: [
+      plannedCommandAlternatives(['widget', 'widget add']),
+      /widget\s+add/i,
+      /label=["'](?:Add Widget|Edit|Done)["']/i,
+    ],
   }),
   makeCase({
     id: 'replay-maintenance-update',