test(verify): autonomous capture-mode toggle verifier - #167

Description

Adds the autonomous installed-APK verifier for the capture-mode toggle — the top-right
CAMERA | VIDEO selector on the viewfinder — and writes "mark INVENTORY when it passes" into the
create-verifier process, which until now happened by hand.

No shipped app code changes: verifier plus skill docs only.

The verifier

verify-openloop/helpers/photo_mode_loop.py drives the real APK on an emulator through
Video → Camera → Video, and asserts each state out of a single uiautomator dump, two ways at
once:

The shutter label is the ability being switched — whether the next tap records a clip or takes
a picture. The checked flag is the indicator — which segment carries the lime fill, and the
reason the single toggling icon became a segmented control at all (issue #126: with one icon, the
only state indicator sat under the fingertip at the exact moment it changed). Half a flip fails: a
highlight that moves while the shutter still records, or a shutter that flips while the pill stays
put.

Both reads come off one dump on purpose. A dump costs seconds on this screen, so polling for the
label and then re-dumping for the flag is a race the harness loses intermittently — and the two
halves disagreeing is exactly the bug being looked for.

Nothing is captured: no still, no clip, no gallery write, no DataStore reset. The loop runs in
about 25 s.

Where selection state actually lives

verify_common.py nodes now carry checkable / checked (in both parsers). Finding the flag
meant reading the real hierarchy rather than guessing at it: Compose puts the selectable on a
full-height wrapper around the visible pill, and only that wrapper reports checked. The label
leaf and the android.widget.RadioButton leaf inside it both report checked="false" whichever
segment is selected — so a verifier that reached for the obvious-looking RadioButton node would
have asserted a constant and passed forever.

uiautomator returns a flat node list, so the loop finds the wrapper geometrically: the smallest
checkable node whose bounds contain the label's. On the shipped screen that resolves to exactly two
nodes, [523,175][786,301] and [786,175][1027,301].

Evidence, and proof the assertions discriminate

Run on emulator-5584 (Pixel_8, API 34):

• PASS serial=emulator-5584 selector=Camera|Video video->photo->video took=26s from .claude,
  and 21 s from the synced .cursor copy — which is the one run-verification-loops.py globs.
• Re-parsing the three saved XMLs afterwards, each state matches its own mode and fails the
  other. A verifier that passes while asserting nothing is the failure mode worth ruling out
  explicitly, not assuming away.
• Both parsers (etree and the regex fallback) agree node-for-node on a real dump, so the fallback
  path carries the new fields too.
• Evidence per run: video-baseline, camera-mode, video-restored as XML + PNG.

Sweep gate 5b picked the loop up with no registry change, since run-verification-loops.py globs
*_loop.py: PASS loops=onboarding,photo_mode,record_clip.

INVENTORY, and making that a step

INVENTORY.md gains a row for the capture-mode toggle as automated 2026-09-01. "Photo stills
mode" deliberately stays mapped — this loop drives the toggle, not the capture, and the table's
own legend makes automated a claim about what a loop drives. Overstating it there is how a green
table starts lying.

Marking that row was ad-hoc (#163 did it by hand as well), so create-verifier/SKILL.md now names
it as step 6 — only after the loop has been seen passing — with the harness sync moved to the end
so it picks up the INVENTORY edit too.

Manual QA / not verified here

• Taking the actual still is still a manual control.ps1 recipe; this loop stops at the toggle.
• The toggle is hidden while recording and mid-booth (CameraScreen), and setCaptureMode refuses
  a switch mid-recording. Neither guard is driven here — that is record-clip / photo-booth
  territory.
• One emulator (Pixel_8, API 34). Not run on a Fold or a physical device.

Checklist

• 🧪 My changes have been verified locally and work as expected.

• 🔍 I have performed a self-review of my own code.

• ✍️ I have commented my code, particularly in hard-to-understand areas.

• 📖 My changes generate no compile warnings or errors (allWarningsAsErrors is on).

• 🧹 Pre-PR sweep GREEN on the final commit (.\scripts\pre-pr-sweep.ps1 →
  build/sweep-receipt.json), run as -SkipInspectCode -SkipConnected. Two gates are therefore
  NOT in the receipt, and both need the owner before merge:

  • Inspect Code (Engine 2): SKIPPED — needs the IDE run with the "OpenLoop Tracked" scope,
    which is not available headless on this machine.
  • Instrumented tests + the verification loops (gates 5 / 5b): SKIPPED — -SkipConnected
    skips both. This is a host limit, not a code one: this machine cannot hold a full sweep and the
    AVD at once (the Gradle daemon reaches ~3 GB and does not give it back), and the owner chose
    this over freeing memory. What was actually observed at this same commit, across three full
    sweep attempts, is below.

  Attempt	Gate 5b (loops)	Gate 5 (instrumented)
  1	PASS loops=onboarding,photo_mode,record_clip	red before running — a phantom adb device appeared, ANDROID_SERIAL unset
  2	PASS loops=onboarding,photo_mode,record_clip	121 tests, 1 failure
  3	PASS loops=onboarding,photo_mode,record_clip	died at 78 tests

  Neither gate-5 red is an assertion failing. Attempt 2's was
  BoomerangEditorScreenTest#selectedLook_reportsSelectedSemantics timing out at exactly 45 s in
  ActivityScenarioRule.after waiting for DESTROYED — teardown, not the one-line test body — with
  2.6 GB free; stopping the Gradle daemons took the host to 6.0 GB and that class then ran 25/25
  green in isolation. Attempt 3 died with DeadSystemRuntimeException: DeadSystemException, the
  emulator's own system_server going down with 1.1 GB free. connectedDebugAndroidTest still
  needs one clean run on a machine with headroom before merge.

• 🏪 Play-facing docs aligned: no permission, data-collection, storage or user-facing change
  — verifier and repo tooling only.

• 🧹 The git branch is clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01249no1TrwKLR1aorGqaf7C