android_ui: never check or dismiss the keyboard on a virtual display

keyboard_is_up read the phone-wide mInputShown, but Android draws the IME
on display 0 even when its target window is on a virtual display, so it can
never cover a VD tap. With two sessions on one phone, the other session's
keyboard on the real screen made dismiss_keyboard press BACK on this
display, popping the route under test and then closing the app.

Prunes the oldest evidence artifact to stay under the 60-file cap.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KeVXsWTuwjmZK5uvbw6vqF

docs/superpowers/evidence/android-ui-vd-keyboard-2026-09-27.json

@@ -0,0 +1,40 @@
+{
+  "intent": "Two Claude sessions sharing one phone must not disturb each other: a keyboard open on the real screen must never make android_ui press BACK on another session's virtual display.",
+  "scope": [
+    "python_pkg/android_ui/_text_entry.py (keyboard_is_up only)",
+    "Non-goal: no change to real-screen (display 0 / no display) behaviour, which still checks and dismisses the IME"
+  ],
+  "changes": [
+    "keyboard_is_up returns False when the driver targets a virtual display (display not None and not 0), so dismiss_keyboard sends no ESCAPE/BACK there",
+    "Test: a VD driver with the phone-wide keyboard shown runs no adb command at all"
+  ],
+  "verification": [
+    {
+      "command": "adb shell dumpsys input_method (live phone, 2026-09-27)",
+      "result": "pass",
+      "evidence": "every IME binding reads imeDisplayId=0, including one whose targetWin was com.kuhy.workout_app.sandbox on targetDisplayId=13: the IME is drawn on display 0 for VD windows"
+    },
+    {
+      "command": "python -m python_pkg.android_ui --display self tap ... (before the fix, while another session typed on the real screen)",
+      "result": "fail",
+      "evidence": "'the soft keyboard is still covering the screen after ESCAPE and BACK'; the BACK popped the workout route, a second run closed the app on the VD"
+    },
+    {
+      "command": "same taps after the fix",
+      "result": "pass",
+      "evidence": "tapped 'Dumbbell Lunge progression settings' / 'Warmup set' / 'Wins needed down' on the VD with the other session still active"
+    },
+    {
+      "command": "pytest python_pkg/android_ui --cov=python_pkg/android_ui --cov-branch",
+      "result": "pass",
+      "evidence": "87 passed; _text_entry.py 100% lines and branches"
+    }
+  ],
+  "risks": [
+    "A future Android build that draws the IME on the VD itself would make taps under it land on keys; the phone would have to show imeDisplayId=<vd> for that, which it does not today"
+  ],
+  "rollback": [
+    "git revert this commit",
+    "After rollback, VD taps work again only while no other session has a keyboard open on the real screen"
+  ]
+}

docs/superpowers/evidence/nvidia-config-allowlist-2026-08-23.json

@@ -1,33 +0,0 @@
-{
-  "intent": "Remove nvidia_config.sh's coverage exemption, which was reinstated when an earlier commit's allowlist deletion did not survive alongside the seam restore.",
-  "scope": [
-    "meta/shell-coverage-allowlist.txt (one line removed)",
-    "Non-goal: any code change; the lib and its tests are already correct"
-  ],
-  "changes": [
-    "Deleted the fixes/lib/nvidia_config.sh line from the allowlist, taking it from 85 to 84 entries"
-  ],
-  "verification": [
-    {
-      "command": "shell_coverage_jail.sh with the gate's own flags (--bind /etc --bind /usr/local/bin --bind /var/lib), in a detached worktree",
-      "result": "pass",
-      "evidence": "nvidia_config.sh: 76/76 lines = 100.00%."
-    },
-    {
-      "command": "The same measurement in the main working tree",
-      "result": "fail",
-      "evidence": "'the subject modified the working tree; a write escaped the jail', listing python_pkg/endurain_import files another agent was editing concurrently. This is a false positive from the jail's tree fingerprint, which is why the real measurement was taken in an isolated worktree."
-    },
-    {
-      "command": "grep for the four seams in nvidia_config.sh",
-      "result": "pass",
-      "evidence": "All four present at lines 26, 27, 60 and 85. An earlier `grep -c 'XORG_CONF:-'` returning 1 counted matching LINES, not seams, and briefly looked like a regression."
-    }
-  ],
-  "risks": [
-    "The coverage gate cannot be trusted while another agent is writing to the same working tree; any measurement must be taken in a detached worktree or it may abort on their churn."
-  ],
-  "rollback": [
-    "Re-add the nvidia_config.sh line to meta/shell-coverage-allowlist.txt"
-  ]
-}

python_pkg/android_ui/_text_entry.py

@@ -28,6 +28,7 @@ class TextEntryMixin(ABC):
     """The text-entry half of :class:`~python_pkg.android_ui.driver.AndroidUi`."""
 
     _settle: float
+    _display: int | None
 
     @abstractmethod
     def _run(self, *args: str, timeout: float = 30.0) -> str:
@@ -137,7 +138,17 @@ class TextEntryMixin(ABC):
         Without this, a caller cannot tell "the button is absent" from "the
         button is behind the keyboard", and the accessibility tree reports the
         button's laid-out position either way.
+
+        Always False on a virtual display: Android draws the IME on the
+        default display (``imeDisplayId=0``) even when its target window is
+        on a VD, so it can never cover a VD tap -- while ``mInputShown`` is
+        phone-wide. Checking it there meant another session typing on the
+        real screen made :meth:`dismiss_keyboard` press BACK on *this*
+        display, popping the route under test (2026-09-27, two sessions on
+        one phone).
         """
+        if self._display not in (None, 0):
+            return False
         out = self._run("shell", "dumpsys", "input_method")
         match = re.search(r"mInputShown=(\w+)", out)
         return match is not None and match.group(1) == "true"

python_pkg/android_ui/tests/test_text_entry.py

@@ -98,6 +98,22 @@ class TestKeyboard:
         ui.device.keyboard_shown = False
         assert not ui.ui.keyboard_is_up()
 
+    def test_never_up_on_a_virtual_display(
+        self, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        # The IME draws on display 0 even for a VD's window, so it cannot
+        # cover a VD tap -- but mInputShown is phone-wide. Reading it there
+        # let another session's keyboard make dismiss_keyboard press BACK on
+        # this display and pop the route under test (2026-09-27).
+        device = FakeDevice()
+        device.keyboard_shown = True
+        vd = drv.AndroidUi(settle_seconds=0.0, display=7)
+        monkeypatch.setattr(vd, "_run", device.run)
+        monkeypatch.setattr(f"{MOD}.time.sleep", lambda _s: None)
+        assert not vd.keyboard_is_up()
+        vd.dismiss_keyboard()
+        assert device.calls == [], "no dumpsys, and above all no BACK"
+
     def test_dismiss_is_a_no_op_when_already_closed(self, ui: Harness) -> None:
         ui.ui.dismiss_keyboard()
         assert not [