Evidence reproduction: mouse back and forward in desktop Browser iframes

Environment for both captures: Linux X11 under headless-desktop/Xvfb;
Electron 44.10.1, --no-sandbox --disable-gpu; 1600x900 app window at (0,0),
light theme. Videos capture the upper 1600x760 of that window; composer
excluded to avoid publishing operational model metadata. Header and footer
are editorial labels, not application UI. The right panel is masked during
fixture setup/reset to exclude machine identifiers that can briefly appear. Native mouse input uses XTest via
xdotool buttons 8 and 9, with distinct press and release. No synthetic DOM
mouse events and no provider turn. This is virtual native input, not a
physical mouse attached to the VM.

Build each checkout separately: vp i; fetch Electron with its install.js;
vp run --filter @t3tools/desktop build:bundle.
Launch each built apps/desktop/dist-electron/main.cjs with its own bundled
Electron in headless-desktop. Set T3CODE_DEV_SEED=0, T3CODE_HOME to that
checkout's .t3, XDG_CONFIG_HOME to that checkout's .t3/config,
T3CODE_DISABLE_AUTO_UPDATE=true, distinct T3CODE_PORT and
--remote-debugging-port. Retain DISPLAY and XAUTHORITY from the launcher.
Both servers start with fresh data; no real database is copied.

Create the same empty synthetic project through:
node apps/server/dist/bin.mjs project add --base-dir .t3 --title 'Navigation proof' <empty-fixture-directory>
Use seed.cjs CHECKOUT CDP_PORT PROJECT_ID on each side to create Thread A
and Thread B through the real authenticated orchestration RPC. It never
starts an agent. Complete onboarding with imports deselected if needed.
Set both windows to 1600x900 with xdotool windowsize/windowmove.
Start fixture-server.cjs on loopback port 18511.
Run evidence-driver.cjs CHECKOUT CDP_PORT DISPLAY XAUTHORITY OUTPUT_ROOT SIDE.
Render labels with render-evidence.py RUN_DIR Before|After SHORT_SHA.

Scripted interaction on both sides:
1. Open A then B, open Browser p1, click its link to p2. Both pages have a
   same-origin iframe so the same pointer coordinate stays over an iframe
   when head goes back to p1.
2. Move inside the iframe; hold Back (button 8), release. Hold Forward
   (button 9) at the same coordinate, release.
3. Open A then B, replace Browser with the boundary fixture. Explicitly
   call the public Electron webview clearHistory() method: a newly created
   app tab retained an initial blank history entry, so new tab alone did
   not create a real boundary. Verify canGoBack=false, canGoForward=false.
4. Hold/release Back, then hold/release Forward at the same coordinate.
5. Each endpoint still was captured only after successive raw pixel
   screenshots 500 ms apart matched. Full interaction recordings are the
   primary evidence. On base, Back routes the host to A and hides B's
   Browser; Forward restores B, revealing the unchanged p2/boundary page.

Scope: same-origin iframe history, release timing, and a deliberately
established empty history boundary. Cross-site iframes, other platforms,
physical hardware, guest recording cursor behavior, and non-thumb buttons
are not verified by these captures.

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