[wingleeio]

Stacked on #551, which makes full-rate frames plain tiles. I'll retarget this to main when #551 merges.

What was wrong. At 120 Hz the display link fires about 2.5 ms before a refresh point. A frame takes about 0.8 ms to render and wait for the GPU, so it's ready about 1.5 ms before the point, and the window server shows it on the refresh after. About one frame in twenty is ready earlier, 1.7–3.2 ms before the point, and shows on that refresh instead. Two frames then land in one refresh and the next shows nothing new. That's why tiles alone still read a 17.7 ms p99 interval.

Measured with host-clock stamps of each tile commit against the display link's output grid (93% of commits at 1.4–1.6 ms before a point). The same phase holds on the AV path, where the AV renderer's own copyDisplayedPixelBuffer showed it taking 99.5% of frames: the loss is in which composite a frame reaches.

The change. align_wait: a frame ready 1.85–4.5 ms before a point waits until 1.6 ms before it, in one short sleep. About 5% of frames wait, each under 1.5 ms. Commits then all land 1.0–1.75 ms before a point, and handover intervals run 7.9–8.8 ms (p1–p99).

Applies only when all of these hold:

• the frame is the one a display refresh asked for (refresh_frame): input's frames are drawn and shown at once, as before;
• the refresh is under 12 ms: 60 Hz, and the latch's steady app-timer frames (refresh_grid), are untouched;
• no resize is settling (500 ms): those frames show at once with the window's new size.

Why not a hold on every frame. Holding every frame past the point gave p99 12.7–12.8 ms, but on the AV layer it cost +0.41 points of CPU (two wakeups and a split frame every refresh). Sleeping once didn't change that, and neither did timed AV presentation (PTS on the refresh grid, p99 only −1.4 ms). Only the early tail needs moving.

Measurements. One quiet window, Zeron hidden, one binary (tiles = #551's path), 8 interleaved ScreenCaptureKit captures each (scripts/framepace.swift), and 4 interleaved suite sampler runs (scripts/macsample.swift) each. The desktop was busy, so compare within the run.

variant	p99 median [range]	p95	fps	CPU (ps)
main (AV layer)	20.55 ms [18.9–21.2]	16.81	103.3	5.24%
tiles (#551)	17.69 ms [17.4–17.9]	12.78	107.1	4.47%
tiles + this	13.63 ms [13.3–15.0]	12.48	108.4	4.47%
tiles + hold every frame (diagnostic)	12.79 ms [12.7–13.1]	11.95	107.4	4.53%
GPUI Fast	13.62 ms [13.2–17.0]	12.70	107.5	6.24%

No variant had long intervals (>25 ms).

variant	CPU, one core	GPU power	WindowServer CPU / GPU	footprint
main (AV layer)	4.84%	0.1551 W	25.2% / 11.19%	253.3 MB
tiles (#551)	4.03%	0.1557 W	23.8% / 11.19%	241.8 MB
tiles + this	4.06%	0.1544 W	24.5% / 11.19%	241.7 MB
GPUI Fast	6.26%	0.1831 W	27.1% / 6.40%	253.4 MB

Offline (10 s runs, /usr/bin/time -l, 3 alternating rounds), against tiles: involuntary context switches 10.3K vs 10.3K, instructions 2.25 G vs 2.25 G per run.

Checks:

• only_early_frames_at_a_fast_refresh_wait_for_the_others_point covers align_wait: the early tail waits to ALIGN_AT, the main cluster and late frames don't, frames a refresh early don't, and 60 Hz never waits.
• Windowed animate resized four times at full rate with ZGX_DAMAGE_CHECK=1: no stale pixels, a full repaint at each resize.
• No CoreMedia threads at full rate with tiles (sample: 0, against 5 with ZGX_PRESENT_VIDEO=1).
• fmt, clippy 1.99 -D warnings, and the zgx-desktop tests (176 passed).

The input-latency probe (macos_latency_probe, typing) is in the shared verification window with #551. Typing frames aren't refresh frames, so they never wait.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[github-actions]

Instruction counts (cachegrind, exact) against d1212406; the gate fails above +1%.

scene	base	head	change
heavy-both	1,415.7 M	1,415.7 M	-0.00%
heavy-stream	1,359.3 M	1,359.3 M	-0.00%
heavy-scroll	80.4 M	80.4 M	-0.00%
lab-stream	462.1 M	462.2 M	+0.03%
lab-scroll	70.3 M	70.3 M	+0.01%
animate	397.6 M	397.6 M	+0.00%
micro-state	59.7 M	59.7 M	-0.00%
micro-vscroll-fixed	48.8 M	48.8 M	-0.01%
micro-vscroll-variable	89.1 M	89.0 M	-0.05%
micro-state-luau	83.0 M	83.0 M	+0.01%
micro-vscroll-fixed-luau	66.3 M	66.3 M	-0.01%
micro-vscroll-variable-luau	134.6 M	134.6 M	+0.03%
mount	327.0 M	327.1 M	+0.04%
script-update	1,700.1 M	1,700.0 M	-0.01%
anime	4,791.6 M	4,791.7 M	+0.00%
script-layout	4,001.4 M	4,001.9 M	+0.01%
list-state	894.4 M	893.7 M	-0.07%
style	2,146.7 M	2,146.7 M	+0.00%
widgets	2,685.2 M	2,685.2 M	+0.00%
app-mail	318.5 M	318.6 M	+0.04%
app-chat	53.2 M	53.2 M	+0.07%
app-dashboard	189.8 M	189.8 M	-0.01%
app-dashboard-sort	126.5 M	126.5 M	+0.01%
app-sketch	48.5 M	48.5 M	+0.12%
app-code	100.0 M	100.0 M	+0.03%
app-code-search	87.6 M	87.7 M	+0.04%