UI bug audit closeout evidence

Date: 2026-08-29

Source commit: d0ce8b885e365893ccbce0135c596cf5b9569f61

This directory records the real built-app acceptance for the Check out all branches as worktrees dialog after WP4. The run used a disposable repository with four local candidates, two remote-only candidates, and main already checked out. The fixture carried no real account, provider, repository, or user data.

Build provenance

Field	Value
Required build	`npx --no-install cross-env RELEASE_CHANNEL=development DESKTOP_SKIP_PACKAGE=1 yarn build:prod`
Required build result	Exit 0, `client_ok: true`, no timeout, 354.52 seconds
Capture build	`npx --no-install cross-env RELEASE_CHANNEL=production DESKTOP_SKIP_PACKAGE=1 yarn build:prod`
Capture build result	Exit 0, `client_ok: true`, no timeout, 388.50 seconds
Capture `out/main.js`	4,630,229 bytes, SHA-256 `ffebecdf23afbca48d5a6e0d162e5426c4275d6ec1d57ab02f78a364ca28dffa`
Electron executable	226,577,920 bytes, SHA-256 `082d352efc6a9f5882354ee4096ae0b40b78bc6c8e52fc5084f3df9254c613ff`
Fixture commit	`8b4bf1647b2f445fc5b815d4d4479c954912b33d`
Theme and language	Light, English

The required development-channel build compiles __DEV__=true, so it programmatically installs two DevTools extensions. Two candidate launches were rejected before interaction because /json/list contained one page plus those two service workers. The capture build uses the same source with the production channel, which compiles out the development-only extension installer. Each accepted run then exposed exactly one page target with the exact expected file URL and a loopback WebSocket.

Accepted captures

File	State	CSS viewport	Physical pixels	Bytes	SHA-256
`checkout-branches-unfiltered-100pct.png`	Six candidates selected, unfiltered actions	1240×725	1240×725	70,935	`383aabc8510ca85779a8146aec2822343057092e908fccd6dbdbfb2ce9ea4fad`
`checkout-branches-filtered-100pct.png`	`feature` filter, shown rows cleared	1240×725	1240×725	60,396	`f8c71a89b1c6d3165ca8f0d276beb8612bec4e831f87ff17baacb79999e13b68`
`checkout-branches-unfiltered-125pct.png`	Six candidates selected, unfiltered actions	1240×726	1550×907	90,415	`5f7bd1fcd08aaf8ebfbbd10e9bc1bf341a36bc40f704ad7394127a2b09cdae2c`
`checkout-branches-filtered-125pct.png`	`feature` filter, shown rows cleared	1240×726	1550×907	78,331	`0c4951637a1fe3da60729fb88eecb8827da62c02ef4d3c72e32b5da9a42e449a`

At 125%, a 725 CSS-pixel height maps to 906.25 physical pixels. The frameless Windows client quantizes that to 907 physical pixels and reports 726 CSS pixels. This one-pixel expansion is recorded rather than hidden. Width remains exactly 1240 CSS pixels, DPR is exactly 1.25, and renderer zoom is exactly 1.

100% scale

All six branch candidates selected at 100 percent display scale

Only the two feature branches shown and cleared at 100 percent display scale

125% scale

All six branch candidates selected at 125 percent display scale

Only the two feature branches shown and cleared at 125 percent display scale

Behavioral proof

The verifier exercised the real dialog state rather than injecting a mock:

With no filter, Select none changed the footer to 0 of 6 branches selected.
Select all restored 6 of 6 branches selected.
Typing feature left exactly feature/alpha and feature/beta visible and changed the buttons to Select shown and Deselect shown.
Deselect shown cleared those two visible rows while retaining all four hidden selections, proven by 4 of 6 branches selected.
Select shown restored the full 6 of 6 branches selected state.

At both scales, the virtual-list slot is exactly 48 CSS pixels tall. The branch name and destination path occupy separate lines inside that slot. The native checkbox is 18×18 CSS pixels, sits in .branch-worktree-checkbox, and carries a branch-specific accessible name such as Check out feature/alpha as a worktree. The document, dialog, dialog content, and list have no horizontal overflow, and the dialog remains inside the viewport.

The capture lifecycle used named off-screen Win32 desktops. Launch, physical dimension checks, text entry, screenshots, process termination, and desktop closure went through the cheap Lowlevel route. The renderer's direct page CDP target supplied exact measurements. Two Chromium overlay operations used the app-native hook after the background HWND route proved unable to activate that overlay reliably. The public machine-readable details are in receipt.json.

The clone dialog was also inspected at 1240×725 against design/screenshots/07-clone.png. The checked-in reference uses a historical inline repository sheet, while the frozen production shell intentionally uses the existing modal. Within that boundary, the History and Worktree option cards match as sibling surfaces, their label and description hierarchy is intact, and the shared checkbox remains accessible. No retired shell file or shell wiring was restored for visual similarity.

Baseline boundary

The broad unit and lint commands remain red for pre-existing source-contract drift. The task introduced no new failure:

main: 1,065 of 1,065 files, 9,138 tests reported, exit 1;
task: 1,065 of 1,065 files, 9,141 tests reported, exit 1;
the task's failures map to 55 unchanged test files;
rerunning those 55 files on main reproduced the same 55 failing files, the same 91 failure locations apart from one diagnostic line offset, and the same 151 normalized failure titles; and
changed-path ESLint reports five existing diagnostics on the task and the same five plus one obsolete wrapping-label diagnostic on main.

The source type check, design-app tests, and all focused WP suites pass. Full counts and exact commands remain in the task handoff and rolling Discussion.

Cleanup