feat: support classic Stream Deck models, starting with the MK.2

[Micha0827]

What this does

Adds support for Stream Deck models beyond the Neo, with the MK.2 as the concrete target. Instead of assuming a Neo, the device layer derives everything it needs from the control definitions the connected device already reports through @elgato-stream-deck/node:

• drawable keys — buttons whose feedbackType is lcd (those carry a pixelSize)
• page controls — buttons without a display, which is exactly what the Neo touch points are
• the InfoBar surface — the lcd-segment control, if the device has one

Nothing is keyed off a model name, so the Original, Original V2, MK.2, MK.2 Scissor, XL, Mini, and the 6/15/32 modules fall out of the same derivation. Devices that report no drawable keys, such as the Pedal, are skipped rather than opened.

The InfoBar on decks without an LCD segment

This was the one part that needed a design decision. My first attempt rendered the existing 248×58 strip across a run of keys and sliced it into tiles — but the physical gaps between keys cut words in half (ALL AG|ENTS, OPE|N), which looked broken on hardware.

What is in the PR instead: a run of adjacent infobar keys in one row each render a self-contained block of the current page. The first names the page, the rest carry its metrics. Tapping any tile advances the rotation, replacing what the touch points do on the Neo. On the 15-key default this maps cleanly — the four pages each have a heading plus up to three metrics, and the bottom row has exactly five keys for four tiles plus the page indicator.

docs/images/15-key-preview.png in this branch shows the result.

Backwards compatibility

This mattered more to me than the new feature, so it is verified rather than asserted:

• The Neo render path is byte-identical. I rendered npm run preview before and after the change and compared checksums — same hash. Key modules keep a fixed 96×96 design canvas and are scaled to whatever pixel size the device reports, so the Neo simply renders at 1:1 as before.
• All 46 existing tests pass unmodified. Every new parameter is optional and defaults to Neo behaviour: renderDeckBuffers(…, profile = NEO_PROFILE), actionForControl(…, profile = NEO_PROFILE), and the key renderers default to a 96×96 size.
• Existing config.json files keep working. An 8-key layout stays valid; if it does not match the connected device, the recommended layout for that device is used instead.
• Collectors, state.ts, and dashboard.ts are untouched — this is purely device, render, and input.

18 new tests cover profile derivation from synthetic control definitions, layout fitting, InfoBar spans, 15 keys rendered at 72×72 with no LCD buffer, per-tile content, and paging from any tile. npm run check is green at 64 tests.

Tested on hardware

Run on an actual Stream Deck MK.2 on macOS: device detection, all 15 keys at 72×72, the InfoBar tiles, and paging through all four pages from any tile. One thing that only showed up on the device: at 72px the scaled-down canvas was readable but tight, so where the smallest hint line is dropped the freed space now goes to the rows that carry information (last commit).

A note on scope

I am aware this widens what a project called Neo Agent Deck does, and that this is your call rather than a technical question. The name is untouched, the Neo behaviour is unchanged, and all device knowledge lives in one new file (src/device.ts), so it should be straightforward to reject or reshape. Happy to split this into smaller PRs, drop the typography commit, or take a different approach to the InfoBar if you had something else in mind.

Commits

1. feat(device) — derive deck geometry from control definitions
2. feat(render) — scale modules to the device and put the InfoBar on keys
3. feat(deck) — drive any deck that reports drawable keys
4. feat(scripts) — npm run preview:15key and multi-device doctor
5. test — cover the 15-key layout
6. docs — document support for classic models
7. fix(render) — enlarge type on keys too small for the full layout

Each commit builds and tests green on its own.