detached: a frontend that has not finished reading is not a screen that differs

`detached session: input from a frontend reaches the core and comes back as a
diff` failed about three runs in five, always the same way: cell 122, the core
holding `g` and the frontend a space. Cell 122 on a 60-wide grid is row 2
column 2, and row 2 is the pane shell's first line — the `g` is the front of
`goblin@pardes ...`, the prompt bash printed into the pty at whatever moment it
felt like. Nothing in the transport was wrong. The frame carrying that prompt
was sent, arrived, and was sitting unread in the client's socket: instrumenting
the mismatch printed `drained 1 more queued messages` and then
`same after drain: true`.

`pumpUntil` returns the instant it decodes the message it was asked for and
abandons the rest of the burst, while every `pump` presents one frame. One
`c.wait(5)` that times out — and under a loaded test binary one does — buys a
second pump before the first frame is read, and from there the client is one
frame behind for the rest of the test. Harmless while the only thing in that
frame is nothing; a cell that differs the moment a forked shell writes its
prompt.

So the assertion was comparing the core's NOW against the frontend's THEN, and
the fix is the one this file already made for two frontends: converge.
`pumpUntilShowsCore` is `pumpUntilSameScreen` with one client instead of two —
pump, wait, drain EVERYTHING, compare — checked before the first pump so a test
already in sync spends nothing, and handing its last comparison to
`expectSameScreen` so a transport that genuinely drops a cell still fails by
naming it rather than by timing out.

Both single-frontend screen assertions take it; the two-frontend one already
had its own. 8/8 clean gui runs against 3-failures-in-5 before, and the tty
suite's 457 unchanged.

0 files changed, 0 insertions, 0 deletions