browser_list answered a running flag on every row, and since 0.53.0 it
could not be anything but true: only open browsers are listed, and one
whose engine has gone is dropped. A field whose value is a property of
the answer invites a reader to branch on it, and the page did, in three
places - the stage filtered the fleet on it, the layout counted it, the
address bar guarded on it - over a case the answer cannot contain.
Gone from the row, from the three readers, from the tool's description
and from the public table. The address bar asks what it always meant:
whether the browser being watched is in the fleet at all, which is the
question that matters when a pinned pane outlives the browser it was
pinned to.
One test now asserts the whole row - id, focused, url, urls - so
a field creeping back is caught rather than tolerated. The page test that
needed a not-running row is now the one that says the page takes the rows
as they come, and names where the expensive case is actually kept away:
the server, which drops a browser whose engine is gone.
Three mutations, three killed: the filter back in the page, the address
bar without its guard, the field back in the answer. Suite 585. Driven
with the real model from this tree: two browsers opened, two screens
drawn, no script error.