Evan neomantra

Request for `ntpicture`: leverage Ghostty 1.4 Kitty graphics additions (`a=f`, `a=a`, `N=1`, Wuffs `o=z`)

Filed against ntpicture after inspecting Ghostty v1.4.0 (src/terminal/kitty/). The prompt below is written to be handed directly to an agent working in nimbleterminal.dev/ntpicture.

Background: What Changed in Ghostty 1.4 (`v1.4.0`)

In-place root-frame editing & sub-rect composition (a=f,r=1,X=1) vs a=T re-transmission (d2ffeb5ba, f3e98fb72, b8222f4a8)

Takeaway

The difference is even bigger than a speedup:

TinyGo 0.42.0 cannot run the gallery at all for more than about seven seconds, while 0.43.0 runs it faster than stock Go.

Setup

I built the ntcharts3d gallery for wasm with the 0.42.0 and 0.43.0 release tarballs using the demo site's flags, plus a third 0.43.0 build with the default collector, and drove each in headless Chrome 155 with WebGPU on Metal. The harness counts Kitty frames in the terminal's read stream and samples heaps. Four tabs, three runs each, 20 s windows. The site's Go 1.27.1 build is included as a reference.

ds4go demos

ds4go `steering` Hello World

ds4go `svgpad` Pelican

ds4go model list

ds4go model download

ds4go validate

Shaders demo: Go vs TinyGo in the browser

Measured 2026-09-29.

TinyGo's frames are 13–18% faster at the median, but its page leaks about one RGBA frame of memory per frame, which makes it unfit for a page left open.

The shaders demo lives in ~/projects/ntcharts/examples/shaders, not in ntcharts3d. The work is uncommitted on a new branch shaders-stage-timings in ~/projects/ntcharts. All tests pass, including the real-GPU test.

How to try the ppc64le fix

The fix lives on a branch of my openresty/luajit2 fork:

It is one commit on top of openresty/luajit2 master (v2.1-20260914). It adds 64-bit versions of the four number-conversion helpers in src/vm_ppc.dasc and reverts the lj_tab.c workaround from luajit2 PR #274, which was papering over the same bug. Nothing outside PPC64 changes.

Everything I have run so far was under QEMU user emulation, so a run on real POWER hardware is the confirmation this needs. Pick whichever of the following fits your setup.

DS4 Playground

Made with ds4go which uses the very interesting ds4 inferencing engine. These are all works in progress.

svgpad
glyphpad
cadpad
steering

	. Item #1 (Same-ID chunked upload deletes the old image on chunk 1 instead of on m=0) — STILL RELEVANT
	• In graphics_exec.zig:1039-1050, loadAndAddImage still explicitly deletes the existing image and its placements
	when the first chunk (m=1) arrives:
	// Reusing a specific image ID deletes the old image and all of its
	// placements when the new transmission begins, not when it completes.
	if (t.image_id > 0) {
	storage.delete(io, alloc, terminal, .{ .id = .{
	.image_id = t.image_id,
	.delete = true,
	} });

	package main

	import (
	"fmt"
	"math/rand"
	"os"
	"strings"
	"time"

	tea "charm.land/bubbletea/v2"