The backlight, over DDC/CI

modules/home/desktop/backlight.nix. Both panels are held at ONE declared brightness and ONE white point, set through DDC/CI instead of through the OSD's buttons.

What made this possible, and what it explains

The August attempt at a real backlight curve was reverted because the OTHER screen of the time, a TV on HDMI, does not speak DDC/CI at all. That screen left the desk in 09/2026, and both panels that replaced the pair answer:

Display 1  card0-DP-1  AUS:XG27ACS:TCLMTF117031
Display 2  card0-DP-2  GSM:LG ULTRAGEAR:307AZQV4Q196

Reading them the first time also settled an older complaint. The blue light filter looked like it was skipping the LG, and the CTM was reaching it the whole time (drm_info proved the matrix was on both CRTCs). The panels were simply starting from different places:

10000 K is not even in the LG's own capabilities string, which lists 5000/6500/7500/8200 plus sRGB. It is an OSD picture mode reporting a value outside the MCCS set. A warm matrix applied on top of a 10000 K panel at full brightness does not look like the same matrix applied on a 6500 K panel at 45%, which is why the filter read as broken on one screen.

Why it needs no root

The /dev/i2c-* nodes are root:root 0660, and the session still reads and writes them, because logind puts an ACL for the seat's user on them. That is the same mechanism /dev/uinput uses for Sunshine. No i2c group, no udev rule of our own, no sudo.

The retry is not defensive programming, it is a known race

The ACL appears when the session takes the seat. A unit that starts before that sees NO display at all, and ddcutil detect returns empty rather than failing, so the run would be a silent no-op. The script retries up to ten times, two seconds apart, and gives up quietly: the same shape of boot race the mouse's DPI had, and the same answer.

The display number is derived, never written down

ddcutil --display N counts its OWN enumeration, which nothing guarantees across a replug. The script runs detect once per attempt, maps DRM connector to that number, and looks the connector up from my.monitors (rule 11). One detect per attempt and not per monitor: it costs about 1.5 s by itself.

The cost, measured

A full run is about 2.9 s: one detect plus two setvcp calls per panel at roughly 650 ms each. That is fine for a login and for an occasional backlight-sync 30 by hand, and it is exactly why the instant control (SUPER+Vol, hyprsunset's gamma) stays where it is. A key held down cannot wait 650 ms per step.

The curve, and where its times come from

The lamp walks the day: 45 in daylight, easing from 18:00, down to 15 in the small hours. Those levels are declared here, one line per moment where the LIGHT changes, and a moment with no line carries the previous one forward.

The TIMES are not declared here, on purpose. They are read from config.services.hyprsunset.settings.profile, the same list that generates hyprsunset's own config, and the timer's OnCalendar entries are generated from it. The day's rhythm exists once (rule 11), so moving "arriving from work" in that list moves the lamp with it. That coupling is the point: it is one life, not two schedules.

What is NOT shared is the level itself. Colour temperature and luminance are two physical quantities with independent curves, and the only reason one number (gamma) used to stand in for both was that the lamp could not be reached.

A tick that changes nothing costs 7 ms, not 2.9 s. The level that was last applied is kept in $XDG_RUNTIME_DIR, and a tick that computes the same value exits before touching i2c. The runtime dir is cleared on boot, which is exactly the behaviour wanted: a fresh session always reapplies.

The table WRAPS past midnight: the awk default is the last row of the day, so a tick at 00:30, before any row matches, gets 23:30's level instead of falling back to daylight.

The number is the knob, and the same number is not the same light

45 on both is where they started, not a calibration: the ASUS is 350 nits typical and the LG 300, so matching the digit is not matching the photons. level in the module is what to turn when they stop matching by eye.