Desktop connection-catalog corruption reproduction

Environment: Linux x86_64, Electron 44.10.1, headless-desktop (Xvfb + D-Bus + keyring), 1100x780 window, light theme. No real provider turns or personal data.

Use separate worktrees and Chromium profiles. Install dependencies with vp i, then start each side with T3CODE_DEV_SEED=0 and T3CODE_DESKTOP_REMOTE_DEBUGGING_PORT set under headless-desktop vp run dev:desktop. Wait for the desktop build. Keep the runner's environment PID, display and Xauthority for launch.py. Unlock the test keyring if its prompt appears. If running concurrently, use separate XDG_CONFIG_HOME values. The desktop watcher can be paused while launching the built app with launch.py; retain and later stop every process you started.

seed.cjs WORKTREE PORT SAMPLE_FOLDER drives onboarding without importing history, adds SAMPLE_FOLDER through Add project > Local folder, enables project folders, saves one synthetic bearer connection through desktopBridge.setConnectionCatalog, and scopes the sidebar to sample-project. SAMPLE_FOLDER must be an existing folder named sample-project. Both seeds save an 808-byte encrypted catalog.

run.cjs WORKTREE PORT stop.js cleanly closes Electron through CDP. Keep the dev runner alive.

prove.cjs WORKTREE PORT DISPLAY XAUTH ENV_PID PROFILE SIDE OUTPUT writes NUL bytes over the stopped app's catalog, keeping its size, launches the built desktop app with --no-sandbox --disable-gpu, records its startup, waits for two screenshots 500ms apart to match, takes a still, and attempts the same project-filter click followed by Escape. The filter is absent on the failing base. Each run creates a timestamped output directory. PROFILE is an isolated XDG_CONFIG_HOME directory. It reads project counts with read-only SQLite and compares preserved corrupt bytes.

Gray rectangles redact provider controls. The same rectangle is applied on both sides; no other app content is changed. Development tools are hidden before recording. Videos show actual Electron windows, including startup. Raw runtime logs are private; runtime-evidence.txt contains only the relevant observed diagnostics and synthetic-data facts.

AI-assisted; verified with the tests and tools listed above.
