Splitting this out of #12345 §7, where it is listed as needing its own project, at @WokoFlipper's request in https://github.com/omacom/omarchy/issues/12345#issuecomment-5727650774. It lands with no translation runtime, no catalog and no locale: it is a schema change to `default/hypr/bindings/*.lua` plus the cheatsheet reader. It is worth doing on its own merits, because today a description cannot be reworded either.

## What a description is today

`o.bind(keys, description, dispatcher, options)` (`default/hypr/helpers.lua:92`) hands the description to five different roles.

1. **Display text** in the cheatsheet, `bin/omarchy-menu-keybindings`.
2. **Join key**, on both sides of the awk join that merges alternative chords for one action.
3. **Input to the ordering rules**: 52 literal regexes match on description text to set precedence.
4. **Round-trip key** after selection: the caller recovers which row was chosen from the text it displayed. #12363 is the general fix for that shape.
5. **A window matcher**, for 4 binds. `command_from` (`default/hypr/helpers.lua:56-72`) passes the description as the first argument to `omarchy-launch-or-focus-webapp` when a bind is `{ webapp = …, focus = true }`. That reaches `omarchy-launch-or-focus:13`, where it is interpolated into a jq regex tested against the window's own `class` and `title`. Those are set by the application, and no locale changes them, so a translated description cannot match: the bind stops focusing the existing window and launches a new one on every press. The four are `WhatsApp`, `Google Messages`, `Google Photos` and `Google Maps`, at `default/hypr/bindings/applications.lua:28-31`. All four are product names a translator would probably leave alone, which is what makes it a trap rather than a bug.

The scale, at `9c5482c5`: `default/hypr/bindings/*.lua` has 193 `o.bind`/`o.bind_toggle` call sites across 6 files, of which 186 pass a literal description, 173 of those distinct; 2 pass none; and 5 build key and description in a loop, which is what makes the rendered list roughly 230 rows. The consequences of editing them are measured in #12345 §7: translating the 186 silently destroys 4 of 4 alternative-chord merges and reorders most of the rendered rows, with no error, no non-zero exit and no failing test. `test/shell.d/keybindings-menu-test.sh` pins 4 descriptions and, incidentally, 1 of the 52 precedence literals.

## The shape that fixes it

A third field: a stable id the machinery keys on, and a description that is only ever displayed.

The migration can be incremental, because the id can default to today's value:

```lua
-- default/hypr/helpers.lua
function o.bind(keys, description, dispatcher, options)
  local opts = options or {}
  opts.id = opts.id or description   -- every existing bind keeps working
  ...
```

Joins, precedence and round-trips key on `id`; the cheatsheet renders `description`. A bind that gains an explicit `id` becomes editable, and a bind that does not behaves exactly as it does today. The same holds for the user's own `~/.config/hypr/bindings.lua`, which enters the same stream.

## The one thing that decides the design, and which I cannot measure from here

The cheatsheet reads `hyprctl binds`, so whatever carries the id has to survive into that output. Two candidates:

- **A bind option.** If Hyprland passes unknown `hl.bind` option keys through to `hyprctl binds -j`, the id rides along and nothing else is needed.
- **A sidecar map** from bind to id, generated next to the lua files and joined by the cheatsheet. It needs no upstream support, but the chord is not a stable key either: it moves with the keymap, and the same action appears under alternative chords, which is the reason the merge exists. So a sidecar has to be keyed on chord plus dispatcher, not chord alone.

The measurement that decides it is one command on a running session: set an extra key in `opts`, then check whether it appears in `hyprctl binds -j | jq '.[0]'`. @WokoFlipper, @sbelcl, if either of you gets there before I reach a session, that is the whole blocker.

## Already in flight

- #12363 gives `omarchy-menu-select` a `--print-index`, so a caller stops recovering an id by cutting up the row it displayed. That is role 4 above.
- #12362 keys the cheatsheet cache on the active collation. Independent of translation, but the same code path: the row order already depends on `LC_COLLATE` today, and a translated description would reorder rows for a second reason.

## Pilot

@WokoFlipper carries 204 translated binding descriptions as an archive that explicitly must not be applied, and runs the merge machinery daily, so that is the largest ready-made label set and the first real consumer. Checking the schema against it before it lands costs less than finding out after. @sbelcl's runtime is independent of this and does not need to land first.