tabs: one window for every agent; inactive tabs take input; background off by defaultHEADmaster

Measured in Brave on 2026-11-14: an inactive tab takes trusted CDP clicks,
key presses, text and mouse moves, and Page.captureScreenshot renders it
live. Only mouseWheel hangs, waiting for a frame. The upstream guard that
refused input to every hidden page was wrong for inactive tabs, and it is
what made a window per session look necessary.

- New sessions' tabs open inactive in the user's last focused window, in
  the session's tab group; a window is created only when none is open.
- New groups start collapsed, unless their tab is the one showing, and
  expand when a tab in them is brought forward. Existing groups keep the
  user's setting.
- Input is refused up front only in a minimized window (untested there).
- chrome_scroll scrolls hidden pages by script instead of the wheel.
- Focus emulation on attach, so blur-sensitive menus stay open.
- Background mode is off by default. Calls still default to background;
  background:false or chrome_tab activate brings a tab forward.
- Agent guidance, tool descriptions and docs say so.

1 files changed, 1 insertions, 2 deletions

diff --git a/test-suite/unit/chrome-command.test.mjs b/test-suite/unit/chrome-command.test.mjs
index dac719a..28bfe8a 100644
--- a/test-suite/unit/chrome-command.test.mjs
+++ b/test-suite/unit/chrome-command.test.mjs
@@ -174,8 +174,7 @@ test("Doctor warns when this session's Pi tab is hidden, and stays quiet when it
   await hidden.run("doctor");
   const report = hidden.notices.at(-1)[0];
   assert.match(report, /⚠ Pi's tab is hidden because its window is covered by another window/);
-  assert.match(report, /Chrome ignores clicks and typing in hidden tabs/);
-  assert.match(report, /\/chrome background off/);
+  assert.match(report, /Clicks, typing and screenshots still work there/);
 
   const none = harness({ send: withStatus({ windowId: null, tabId: null }) });
   await none.run("doctor");