[lippdev]

Summary

Fixes the three controller problems found in a hands-on test.

1. A on a text field opens the keyboard (instead of Y).

• spatial.js returns edit when A lands on a text field, and InputRouter then opens the controller keyboard for the page. This works on every page, including ours (the new tab's search box).
• The preload no longer opens the keyboard on any focusin. That used to pop it up when the D-pad walked over a field or a site used autofocus. It now opens only when the focus follows a real pointer press (mouse, or A in Cursor mode), and this hook also runs on internal pages now.
• The address field opens the keyboard on mousedown instead of on focus, so reaching it with the D-pad doesn't open it on the way past.

2. The keyboard keys didn't move.

• keyboard-close hands the window's focus back to the tab. From the second opening on, the D-pad walked the keys in a document without system focus, so Chromium drew no focus and the keyboard looked dead. openKeyboard now always calls focusChrome().
• The first press lands on the first key, not on the close button. The dialog still takes the initial focus so screen readers announce it.

3. Up from the top of the page reaches the toolbar and the tabs.

• spatial.js returns edge on up when nothing is above and the page is at its top. Shell.enterChrome reveals the bar (in TV mode it stays visible while it has the focus), takes the window's focus and puts it on the address field (new shell:focus-chrome channel).
• In the bar, the D-pad walks the tab strip and the toolbar and A presses (on the address field it opens the keyboard). B, or down with nothing below, returns to the page through the new focus-page command, which is validated.
• Minimize, maximize and close stay off the D-pad, so one A can't close the app.
• If the keyboard is opened from the bar and cancelled, the focus stays in the bar.

Spatial scoring (both bugs were found by the new tests):

• Page: sharing a column or a row weighs more than where the centres sit. The D-pad used to jump over a wide field under a short link.
• Shell: distance is measured edge to edge, so left from the address field reaches reload instead of the tab strip. Whether a control counts as "ahead" is still judged by its centre, so the close button inside a tab card stays reachable. The test that plans paths in the tab grid mirrors this calculation.

docs/CONTROLLER-TEST.md gains two scenarios for the TV session.

Test plan

• npm run typecheck
• npm test (104 passed)
• npm run test:e2e (51 passed, 3 skipped), with two new tests in controller.spec.ts:
  • the field is reached without opening the keyboard; A opens it with the tab holding the focus (the reported case); the D-pad walks the keys; A types into the page; it works the second time too
  • up from the top reaches the address field (no keyboard); A and then B keep the focus in the bar; left reaches reload, up the tabs, down returns to the page
• Screenshots on the real Windows screen: the key, the "+" and the tab are each drawn with the focus ring
• Physical controller on the TV (docs/CONTROLLER-TEST.md scenarios 5 and 6)

🤖 Generated with Claude Code