## What happens
When every output disconnects and reconnects at the same time (switching a KVM away and back), one monitor briefly shows **two bars stacked**, and the other monitor shows **no bar** at first. It settles to one bar per monitor after 5–15 s on this machine.

## Expected
One bar per monitor at all times, and no bar built for a screen Hyprland doesn't have.

## What I recorded
I polled `hyprctl layers -j` / `hyprctl monitors -j` every 0.5 s. The sequence was the same on every switch (every bar belongs to the single omarchy-shell PID):

| t after switch back | HDMI-A-1 (Samsung) | HDMI-A-2 (Philips) |
|---|---|---|
| 0 s | output back, no bar | output back, no bar |
| +1.6 s | 1 `omarchy-bar` at y=0 | none |
| +6.3 s | **2** `omarchy-bar`: y=0 **and y=26** | 1 `omarchy-bar` |
| +10 s | 1 | 1 |

At the same moment the shell logs:

    INFO qt.qpa.wayland: There are no outputs - creating placeholder screen
    WARN: Layershell screen does not correspond to a real screen. Letting the compositor pick.   (x2)

## Timing measurements
I ran a copy of `$OMARCHY_PATH/shell` with `console.warn` timestamps added: in `BarPanel` `Component.onCompleted`/`onDestruction`, on `Quickshell.onScreensChanged`, and in `ModuleSlot` (slot completion plus each Loader's `onLoaded`). Two switch-backs:

| Event | Switch 1 | Switch 2 |
|---|---|---|
| Bar for the placeholder screen (`screen.name == ""`) done | 12:27:20.3 | 12:28:01.2 |
| HDMI-A-1 bar done | 12:27:26.3 (+6.0 s) | 12:28:04.5 (+3.3 s) |
| HDMI-A-2 bar done | 12:27:36.5 (+10.3 s) | 12:28:09.3 (+4.8 s) |

What this shows:
1. **A complete bar is built for Qt's placeholder screen.** `Variants { model: Quickshell.screens }` in `plugins/bar/Bar.qml` has no filter, so while no outputs exist the placeholder gets a full `BarPanel` with every widget. Quickshell can't map it to an output, so Hyprland places it on the first monitor that returns. That's the extra bar.
2. **The bars are built one after another on the GUI thread.** `onScreensChanged` only fires after each `BarPanel` finishes, so HDMI-A-2's bar can't start until HDMI-A-1's is done. That's why the second monitor has no bar at first.
3. **The placeholder build costs a full bar's worth of time** before the real bars even start, and the placeholder bar stays visible for that whole time.

Build cost per bar, averaged over 6 builds (gaps between events, credited to the widget that finished next, so approximate):

| ms / bar | widget |
|---|---|
| 630 | `omarchy.indicators` |
| 481 | `othavi0.agent-bar` |
| 238 | `crmne.hyprmoncfg` |
| 221 | `tmn73.calendar` |
| 216 | `robzolkos.github` |
| 175 | `omarchy.tray` |
| 60–170 | every other widget, including trivial ones (`omarchy.menu` 78, `omarchy.power` 71) |

34 widgets per bar gives 3–10 s per bar on this CPU. The ~60–80 ms that every slot costs, however small the widget, suggests `ModuleSlot` itself is expensive to create.

## When it's visible
How visible this is depends on timing:
- **Intel N150** (i915), 2× 4K HDMI at scale 1.6, 34 bar widgets: visible on every switch.
- **Ryzen 9 6900HX** (Radeon 680M), same monitors, same plugins, same package versions: never seen. The bars presumably build fast enough that the placeholder bar is gone before it can be seen.

## Possible fixes
- Skip screens in the bar's `Variants` (and `DragGhostPanel` / `BarMoveGhostPanel`, which use the same model) that don't match a real Hyprland monitor, e.g. an empty `name`. Or debounce rebuilding until the screen list is stable.
- Longer term: build bars incrementally or asynchronously, and make `ModuleSlot` cheaper to create, so N monitors × M widgets doesn't block the GUI thread for seconds on low-end CPUs.

## System
- omarchy 4.0.4-1, quickshell 0.3.1-1, hyprland 0.56.2-2, aquamarine 0.15.0-2, linux-omarchy 7.2.5-3
- Intel N150 / Alder Lake-N graphics (i915), 4 cores
- Samsung U32R59x + Philips 328E1, both 3840x2160@60, scale 1.6, via a generic 4-port HDMI/USB KVM

## Repro
1. Two monitors through a KVM, on a slower machine.
2. Switch the KVM to another computer, wait ~10 s, switch back.
3. Watch the bars, or poll `hyprctl layers -j`.

Related, but a different symptom: #12554 (Quickshell abort when no outputs remain), #9975 (IpcHandler registered once per bar).

Filed by Claude Opus 5.5 via Claude Code.