Sylpheed/tools/re-capture/ui_draw_capture.sh

#!/usr/bin/env bash
# ONE blocking session: boot -> title screen -> arm the UI draw-order capture.
#
# Unlike skip_intro.sh this deliberately does NOT tap A on the title: the title
# screen IS the subject, and a stray A there sends the guest into the save-data
# probe (which crashed the guest on run 1 of 2026-08-18). It taps A only while a
# movie is playing, and the instant the title is classified it presses F10,
# which arms `log_ui_draws`' 3-frame submission-order dump.
#
# The log lands in the emulator's CWD, so the emulator is launched from $OUT.
#
# A grab that is not full-width is treated as a HARD ERROR rather than
# classified. Run 1 spent 13 minutes acting on a 10-pixel sliver: the screenshot
# wrapper had picked a 10x10 helper window that shares the class "xenia_canary"
# with the real one. Nothing failed loudly; the classifier just returned noise.
#
# Usage: ui_draw_capture.sh [timeout_s] [out_dir]
set -u
export HOME=/sylph-home/re SDL_AUDIODRIVER=dummy DISPLAY=:98
SD="$(cd "$(dirname "$0")" && pwd)"
TIMEOUT="${1:-900}"
OUT="${2:-/sylph-home/re/uicap}"
mkdir -p "$OUT"
rm -f "$OUT"/xenia_re_ui_draws_*.log

alive(){ ps -o pid=,stat= -C xenia_canary 2>/dev/null | awk '$2 !~ /^Z/ {print $1}'; }
wide(){ [ "$(identify -format '%w' "$1" 2>/dev/null || echo 0)" -gt 1000 ]; }

xsetroot -solid black 2>/dev/null || true
# ATTACH=1 watches an emulator that is already up instead of launching one.
if [ "${ATTACH:-0}" != "1" ]; then
  ( cd "$OUT" && nohup run-canary --log_ui_draws=true --ui_draw_capture_frames="${FRAMES:-3}" --ui_draw_capture_max="${MAXDRAWS:-20000}" ${EXTRA_FLAGS:-} >"$OUT/canary.stdout" 2>"$OUT/canary.stderr" & )
  # Grace period before the liveness check: run-canary is a shell that execs the
  # binary, and polling `ps -C xenia_canary` in the first moments reports GONE
  # for a process that is merely not exec'd yet.
  #
  # ⚠️ GRACE is why an ARM=early capture still MISSES BOTH BOOT SPLASHES. They
  # run at roughly 1.2-9.5 s of guest time (`boot-order-and-splash-dwell.md`),
  # and a fixed 8 s wait plus the window poll arms after the publisher has been
  # and gone. Set GRACE=1 to catch them; the liveness check below tolerates it
  # because it polls rather than sampling once.
  sleep "${GRACE:-8}"
fi

until xdotool search --name "Xenia-canary" >/dev/null 2>&1; do
  [ -n "$(alive)" ] || { echo "EMULATOR GONE before the window appeared"; exit 4; }
  sleep 1
done
win="$(xdotool search --name "Xenia-canary" | tail -1)"
echo "WINDOW=$win"
xdotool windowactivate "$win" 2>/dev/null; xdotool windowfocus "$win" 2>/dev/null

# 🔴 NOTAP=1 stops this script tapping A at all. The tap exists to skip the
# attract movie, but its trigger is "the screen changed a lot", which is ALSO
# true while a boot splash is fading. A run measuring the BOOT sequence must set
# it: with the tap on, a 2026-08-29 run classified the publisher splash as a
# movie at t=3 s, tapped through it, and the developer splash never appeared at
# all. The instrument was perturbing the thing it was measuring.
#
# ARM=early presses F10 before the title exists, so a long window contains the
# frame in which the screen is BUILT. Armed at the title instead, a capture only
# ever sees the steady state — and on this title screen the steady state is 11
# draws a frame that never mention a sprite.
if [ "${ARM:-title}" = "early" ]; then
  xdotool windowactivate "$win" 2>/dev/null
  xdotool key --window "$win" F10
  xdotool key F10
  echo "ARMED EARLY at ${SECONDS}s"
fi

deadline=$(( SECONDS + TIMEOUT ))
while [ $SECONDS -lt $deadline ]; do
  rm -f /tmp/u1.png /tmp/u2.png
  screenshot /tmp/u1.png >/dev/null 2>&1; sleep 0.6
  screenshot /tmp/u2.png >/dev/null 2>&1
  [ -s /tmp/u1.png ] && [ -s /tmp/u2.png ] || { echo "SCREENSHOT FAILED at ${SECONDS}s"; exit 5; }
  wide /tmp/u2.png || { echo "GRAB IS NOT THE GAME SURFACE ($(identify -format '%wx%h' /tmp/u2.png)) at ${SECONDS}s"; exit 6; }
  [ -n "$(alive)" ] || { echo "EMULATOR GONE at ${SECONDS}s"; exit 4; }
  d=$(compare -metric RMSE /tmp/u1.png /tmp/u2.png null: 2>&1 | sed 's/ .*//' | cut -d. -f1); d=${d:-0}
  s=$(python3 "$SD/screen_id.py" /tmp/u2.png | awk '{print $1}')
  echo "t=${SECONDS}s screen=$s rmse=$d"
  # TARGET=menu walks one step further: tap A ONCE on the title, then arm on the
  # main menu. One tap, and only from a full-width grab — the run that tapped
  # repeatedly (off a 10-pixel sliver it had classified) crashed the guest in the
  # save-data probe behind that button.
  if [ "${TARGET:-title}" = "menu" ]; then
    if [ "$s" = "title" ] && [ "${tapped:-0}" = "0" ]; then
      echo "TITLE at ${SECONDS}s -> one A"; python3 "$SD/pad.py" tap A 0.25; tapped=1
    elif [ "$s" = "menu" ]; then
      cp /tmp/u2.png "$OUT/menu-before-f10.png"
      xdotool windowactivate "$win" 2>/dev/null
      xdotool key --window "$win" F10
      xdotool key F10
      sleep 3
      screenshot "$OUT/menu-after-f10.png" >/dev/null 2>&1
      if ls "$OUT"/xenia_re_ui_draws_*.log >/dev/null 2>&1; then
        echo "MENU CAPTURED at ${SECONDS}s"; break
      fi
      echo "F10 pressed on the menu; no log yet"
    elif [ "$d" -gt 1500 ] && [ "${tapped:-0}" = "0" ]; then
      # Still in the attract loop. Without this the menu target waits forever:
      # the attract movie returns to the title and away again, and a run that
      # only ever watches never gets there.
      #
      # 🔴 `tapped` guard added 2026-08-29, and it is not cosmetic. The trigger
      # here is "the screen changed a lot", which is ALSO true of the
      # title→menu transition this target has just started. A run tapped A on
      # the title at t=23 s and again at t=27 s on the transition; the guest
      # faulted, and Xenia dumped registers to stdout until the file reached
      # **519 MB**. That is the double-tap into the save-data probe this
      # script's own header warns about, arriving through the movie branch.
      echo "movie (rmse $d) -> skip A"; python3 "$SD/pad.py" tap A 0.25; sleep 3
    fi
    sleep 1
    continue
  fi
  if [ "$s" = "title" ] && [ "${ARM:-title}" = "early" ]; then
    cp /tmp/u2.png "$OUT/title-reached.png"
    echo "TITLE REACHED at ${SECONDS}s (capture was armed early)"
    break
  elif [ "$s" = "title" ]; then
    cp /tmp/u2.png "$OUT/title-before-f10.png"
    xdotool windowactivate "$win" 2>/dev/null
    xdotool key --window "$win" F10
    xdotool key F10
    sleep 2
    screenshot "$OUT/title-after-f10.png" >/dev/null 2>&1
    sleep 2
    if ls "$OUT"/xenia_re_ui_draws_*.log >/dev/null 2>&1; then
      echo "ARMED and writing at ${SECONDS}s"
      # A long window is deliberately NOT waited out here: the caller polls the
      # log, because the interesting frames may be many seconds away.
      [ "${WAIT_DONE:-0}" = "1" ] || break
    fi
    echo "F10 pressed; window still open"
  elif [ "$d" -gt 1500 ] && [ "${NOTAP:-0}" != "1" ]; then
    echo "movie (rmse $d) -> skip A"; python3 "$SD/pad.py" tap A 0.25; sleep 3
  fi
  sleep 1
done

grep -i "UI-CAP" "$OUT/canary.stdout" 2>/dev/null | tail -5
ls -l "$OUT"/xenia_re_ui_draws_*.log 2>/dev/null || echo "NO UI DRAW LOG"
echo "UI DRAW CAPTURE SESSION DONE (emulator left running on purpose)"