A blitter retired for somebody else's code

2026-09-16

The ESP32 panel's original firmware is a blitter: the desktop renders, the device streams pixels over serial. That meets "works when nobody is at a desk" never — these units get carried around, and an unplugged blitter shows nothing. pulse/ is the same panel driven by LVGL instead, so it boots, asks for a network on its own glass, joins, saves to NVS, and fetches trending tokens with no host in the loop. app/ still builds; the blitter is one flash away until pulse/ proves itself on hardware.

None of the three bugs that came out of it would show up in a passing test. The touch controller drops contact mid-press — sensors.cpp found this months ago and settles it with a 40ms delay, but pulse_touch.cpp was written fresh for LVGL and didn't inherit that, so LVGL saw press and release several times a second and no popover survived long enough to draw. The panel's driver wants its column window on an even boundary; an odd one shifts each row a pixel further across than the last, which down a line of text reads exactly as italics — only some labels slant, because whether a box lands on an odd column is luck. And LVGL renders straight into its draw buffer instead of the blitter's streamed framebuffer, so PSRAM's latency gets paid on every glyph; moved to internal SRAM, which the buffer's own comment had already guessed would fix it.

The simulator that made all three findable didn't exist yesterday either. sim/ compiles the actual .ino and .cpp files unmodified against shims for Arduino, the display library, WiFi and NVS, renders into a real framebuffer, and describes every frame in words — each label, its box, whether it escaped its parent — because reading what the screen says is cheaper than reading pixels. It found seven bugs in wifi_setup.cpp on the first run, including a join whose result was never drawn: the unit joined, saved, and sat on "connecting" forever.

Then Ryan looked at the actual device: "the not set up screen is weirdly to the right with a lot of blank space." He was right — every empty state had been laid out as a portfolio reading with one row filled in, so the status word sat in the value column at x=123 with three blank pairs under it and 250px of dead panel below. Fixing it meant writing down what a status screen actually is, rather than copying a reading's layout and deleting most of it. pulse_design.{h,cpp} is that: colour by role, a spacing scale whose largest step is the panel's corner clearance, four named archetypes. Unifying the two copies of the colour palette drifted apart between pulse_ui.cpp and pulse_wifi.cpp changed no pixel — provable now, for the first time.

A unit with nothing saved had no way back into setup short of a 1.4-second hold nothing on screen mentioned, on a touch controller that until yesterday couldn't reliably hold that long. Fixed today too: with no network saved, the whole panel is the button.