683a882f

test: a throughput leg that can see a detached session rendering

a73x 2026-08-21 18:19

Commit message

test: a throughput leg that can see a detached session rendering

The gate could not see the regression the previous commit fixed. Both
existing legs flood `yes`, and a one-character row is nearly free to
render, so they stay flat straight through it. Same box, either side of
the fix:

    solo    (yes, no client)        5ms  vs   6ms
    client  (yes, attached)        48ms  vs  55ms
    repaint (full width, detached) 87ms  vs 479ms

The first two are noise. The third is the bug.

So: a third leg that repaints a full screen with nobody attached — 15MB
of cursor-home plus 23 rows of 79 columns, best-of-5 like the others.
23 rows and not 24 so the screen is repainted rather than scrolled;
scrolling costs memset and cursorScrollAbove, which would drown the cost
being measured.

The ceiling is 200ms, sitting near the good number (87ms, readings
87-90) rather than under the bad one. At 400ms a machine twice this fast
runs the BROKEN build in 240ms and the gate says nothing.

The frame file is built by doubling — 13 cats rather than 8192.

Every arm was watched to fire: the pre-fix binaries trip the ceiling at
497ms, a mutated needle trips the empty-grid guard, and one rep instead
of five trips the reading count. Two of those mutations silently failed
to apply first, and only a diff line count caught it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JDnJ6v15RvwZ2Y3zvfrWpR

test/throughput.sh

@@ @@ -43,6 +43,23 @@ SOLO_MAX_MS=10 @@
+CLIENT_LINES=200000
+CLIENT_WARN_MS=60
+CLIENT_MAX_MS=75
+# Leg 3 repaints a full screen over and over with NOBODY attached. Its
+# own leg because `yes` cannot see what it sees: a one-character row is
+# nearly free to render, so the solo leg above stays flat through a
+# regression that makes full-width output 10x slower. Real output —
+# vim, logs, a build — is full-width, and the daemon renders every row
+# of it to hash it.
+# 2^REPAINT_DOUBLINGS frames of ~1.8KB; 13 is ~15MB.
+REPAINT_DOUBLINGS=13
+# Measured either side of the fix, same box, best-of-5: 87ms with the
+# detached check in place, 479ms without it. The ceiling sits near the
+# good number rather than under the bad one — put it at 400 and a machine
+# twice this fast runs the BROKEN build in 240ms and the gate says
+# nothing. 2.3x over nominal, and the readings span 87-90.
+#
+# 5.5x here against 27x at 200x50: the penalty is per rendered row, and
+# this leg renders 23 of 80 columns rather than 49 of 200.
+REPAINT_MAX_MS=200
+# Repeats per leg; the fastest one is the reading. See the header.
+REPS=5
+    echo "  Not a failure. It is the number that decides whether this feels good to use."
+# ---- leg 3: full-width repaint, nobody attached ------------------------
+# The client from leg 2 has quit, so this is a detached session again.
+#
+# One frame: cursor home, then 23 rows of 79 columns. 23 and not 24 so the
+# screen is REPAINTED rather than scrolled — scrolling is a different cost
+# (memset and cursorScrollAbove) and it would drown the one being measured.
+{
+    printf '\033[H'
+    i=0
+    while [ "$i" -lt 23 ]; do
+        printf '%79s\n' '' | tr ' ' x
+        i=$((i+1))
+    done
+} > "$TMPD/frame"
+# Doubling rather than a loop of appends: 13 cats instead of 8192.
+cp "$TMPD/frame" "$TMPD/big"
+i=0
+while [ "$i" -lt "$REPAINT_DOUBLINGS" ]; do
+    cat "$TMPD/big" "$TMPD/big" > "$TMPD/big2"
+    mv "$TMPD/big2" "$TMPD/big"
+    i=$((i+1))
+done
+REPAINT_BYTES=$(wc -c < "$TMPD/big")
+rm -f "$TMPD/repaint" "$TMPD/repaint.done"
+"$MUXA" send --sock "$SOCK" \
+    "for k in \$(seq $REPS); do A=\$(date +%s%N); cat $TMPD/big; B=\$(date +%s%N); echo \$(( (B-A)/1000000 )) >> $TMPD/repaint; done; echo ok > $TMPD/repaint.done\n" \
+    > /dev/null
+i=0
+while [ ! -s "$TMPD/repaint.done" ] && [ "$i" -lt 120 ]; do sleep 0.5; i=$((i+1)); done
+[ -s "$TMPD/repaint.done" ] || {
+    echo "throughput FAIL: $REPS repaints of $REPAINT_BYTES bytes unfinished after 60s"
+    exit 1
+}
+REPAINT_ALL="$(cat "$TMPD/repaint")"
+REPAINT_MS="$(echo "$REPAINT_ALL" | sort -n | head -1)"
+[ "$(echo "$REPAINT_ALL" | wc -l)" -eq "$REPS" ] || {
+    echo "throughput FAIL: wanted $REPS repaint readings, got: $REPAINT_ALL"
+    exit 1
+}
+# The screen must actually be full of the frame, or a shell that never ran
+# the cat would score a perfect time for doing nothing.
+"$MUXA" capture --sock "$SOCK" | grep -qF 'xxxxxxxx' || {
+    echo "throughput FAIL: the repaint left no output on the grid"
+    "$MUXA" capture --sock "$SOCK"
+    exit 1
+}
+echo "repaint: $REPAINT_BYTES bytes, detached:  ${REPAINT_MS}ms (max ${REPAINT_MAX_MS}ms)  [$(echo "$REPAINT_ALL" | tr '\n' ' ')]"
+[ "$REPAINT_MS" -lt "$REPAINT_MAX_MS" ] || {
+    echo "throughput FAIL: repaint leg took ${REPAINT_MS}ms, ceiling is ${REPAINT_MAX_MS}ms"
+    echo "  A detached session should not be rendering rows at all. Check that"
+    echo "  sendUpdate still asks hasClientsIn BEFORE it diffs."
+    exit 1
+}
+echo "throughput OK"