[NicholasEhsanRoy]

test_a_scripted_session_on_a_terminal_shows_each_event_once_above_a_strip_that_follows_the_status failed with "the live region is gone" after #138 (run 41839710392, [60]; earlier run 41762639938, [100]). #138 made the reader wait for the terminal's end; it could not reproduce the failure naturally. This PR captures the failure, reproduces it, and fixes the cause in ServeDisplay.

Capture

On a failure of its end checks, the session test now prints the screen after the stop (lines numbered), what the terminal received where the strip was first drawn and at its end (control sequences written out, each part at most 4000 characters), and whether the reader saw the terminal's end. A planted failure shows the message.

Reproduction and cause

Conditions: the session test at its three widths beside tests/cli/test_serve_view.py, three pytest runs at once on two xdist workers each, all under taskset on four cores with two busy loops of our own (and on two cores with three, for one arm), on a machine other sessions were also loading. Unchanged code: 2 failures in 225 executions, both at [60], both "the live region is gone", both with the transcript below.

The captured transcript (60 columns) shows the cause. After the load report, the view's draw (its own thread) wrote \r\x1b[2K\x1b[1A\x1b[2K and a 7-line READY region; then serve's ready line (the serve loop's thread) wrote the same 2-line reset, its line and the region. Five lines of the first frame (READY · step 0, Ctrl+C to stop, the table's header) stayed in the scrollback. rich's live display finds the top of its region from the height it was last drawn at under its own lock, but draws the region again only after letting that lock go, so a draw between the two leaves the print moving up by a stale height. The end of the transcript was complete and the reader saw the terminal's end: the clear was not lost. An in-process probe with nothing injected (ServeDisplay on a pseudo-terminal going through serve's load report and ready line, on one contended core) leaves the same ghost in 4 of 5100 trials, the raw bytes showing the same two writes with the same stale reset.

The same transcript shows a second defect of the same kind, on every run: the view kept one live display across the pause around the load report, still holding the spinner's shape, so the first line once the view was back erased the report's last line (derived graph: ...) wherever the building phase takes two lines (60 and 80 columns).

Fix

• Every print through the view's console holds the lock its draws hold (_ordered_console): serve's own lines, the events, the draws, and writes to the redirected standard streams from any thread come one after the other. A line of serve's own can now wait for a draw under way, which stays bounded by the writer's flush wait.
• The view starts a live display of its own each time it comes back, so no reset refers to a region no longer on the screen.

Tests

In tests/cli/test_serve_view.py, each with a control that fails its check on every run:

• A draw made on another thread while the ready line is being printed (between its reset and its region, the window a loaded machine hit) leaves no frame of the strip. Control: a console whose prints are not ordered with the draws. The draw is forced deterministically: the line waits for it to have run or to be held back.
• The first line once the view is back erases nothing printed while it was away (the phase checked to take more than one line). Control: one live display kept across the pause.

Both fail on 60 of 60 runs on the code before this change (10 runs, 3 widths each).

After the fix, the same loop: 0 failures in 180 executions, and the probe 0 ghosts in 3000 trials. At a base rate near 1% neither comparison discriminates on its own (with the unchanged code's rate, no failure in 180 has probability about 0.2, no ghost in 3000 about 0.1). What shows the fix is the pair of tests above, whose interleaving is forced: 60 of 60 failures before, none after, and their controls fail on every run. Stress: 20 of 20 runs passed on tests/cli/test_serve_terminal.py and tests/cli/test_serve_view.py.

Not changed here: RunDisplay (mime run) restarts one live display across its pause too, and prints Ctrl+C's acknowledgement beside rich's own refresh thread; the same class, not this test's failure.

🤖 Generated with Claude Code