docs(log 039): one list, one painter — the LCXL3 driver joins the crew

armada/tasks/039-one-list-one-painter.md

+---
+log: 039
+title: "One list, one painter — the LCXL3 driver joins the crew"
+date: 2026-09-05
+task: "n/a (mid-session convergence pass, PLN's ask)"
+tags: [tooling, rig, systemd, lcxl3, convergence]
+shareable: true
+---
+
+## Cap (what & why)
+
+A new track ("Veridis Quo", algoraoke) played to a silent room twice in one
+night — first the recurring saved-at--inf Ardour faders, then the LCXL3 sitting
+untranslated because `lcxl3-driver.py` only ran when a session remembered to
+start it. When the session's background copy got OOM-reaped, the surface went
+deaf **with every check green**. PLN: *"lets ensure we converge on the tooling,
+all these should be auto setup, discoverable, unified."*
+
+## Manœuvre (how)
+
+- **`tools/lcxl3-driver.service`** — repo-owned like its siblings, symlinked
+  into `~/.config/systemd/user`. Retry-forever hotplug posture (the
+  `StartLimitIntervalSec=0`-in-`[Unit]` lesson from lcxl-leds-watch), follow
+  mode (no `--track`; it tracked rose_rouge on its own within a second),
+  deliberately not niced (it is the surface's nervous system).
+- **`Conflicts=lcxl-leds-watch.service`** — one painter per board, enforced by
+  systemd instead of by memory. Starting either painter stops the other, so
+  plugging a surface generation picks its painter.
+- **`rig_units.py`** — the driver and `perf-tray` join THE list (gear managed
+  perf-tray but the inventory didn't know it — the "list lived in three
+  places" failure, again). New third boot policy **`manual`**: never enabled,
+  never started by `--ensure` or the target, because auto-starting the v2
+  painter would Conflicts=-kill the LCXL3 driver mid-bring-up. `--list --boot`
+  filter added for shell consumers.
+- **`parvagues-gear`** moved from a loose file in `~/.local/bin` into
+  `tools/`, and now **derives** its unit sets from `rig_units.py --list`
+  instead of being the fourth hardcoded copy. Its `autostart-on` used to
+  `enable ALL` — including parvagues-sc at login, against the rig's own
+  policy; it now defers to `--ensure --apply`.
+- **`gig-up.sh leds()`** detects the plugged surface generation: LCXL3 →
+  restart the driver unit; v2 board → the old paint path, now invoked via
+  `python3` so a lost exec bit can't kill it.
+- **63 scripts got their exec bit back in the INDEX** (`git update-index
+  --chmod=+x`) — `core.fileMode=false` means working-tree chmod never reaches
+  the repo, which is why 203/EXEC (log 032) kept recurring.
+
+## Prise (findings / artifacts)
+
+- `rig_units.py --status`: **9 units, 0 problems** — lcxl3-driver active +
+  enabled, lcxl-leds-watch `manual — start by hand for its hardware`.
+- The night's earlier debugging, now structural: LCXL 1 worked plug-and-play
+  because it *natively* spoke the corpus's CC map; the LCXL3 only speaks it
+  through the driver, so the driver is rig infrastructure, not a session tool.
+- Also confirmed on the way: nothing in the toolbox reads **live** Ardour
+  record/transport state — check-mix and friends parse the *saved* session
+  (and say so). A live REC lens is an open want (OSC port 3819, or polling
+  the growing stem files in `interchange/`).
+
+## Sel (the shareable learning)
+
+- **A tool that must run for hardware to work is a unit, not a command.** The
+  test: if the OOM killer reaping it leaves the rig lying ("every check
+  green, surface deaf"), it was infrastructure all along.
+- **Mutual exclusion belongs to the init system.** "Only one painter may own
+  the board" lived in a comment; `Conflicts=` makes it a law that holds even
+  when nobody remembers it.
+- **An inventory earns its keep by being complete, then by having policies.**
+  Two policies (login/on-demand) couldn't express "never auto-start this";
+  the third policy existed the moment two units could kill each other.
+
+## Hameçon (hook)
+
+The surface worked perfectly all night — for the person clicking with a mouse.
+The story of why a fader is a *protocol*, and why the fix was a systemd
+directive from 1992's playbook: `Conflicts=`.