feat: bind key chords from the config file

[clvrobj]

What

Adds an optional "chords" section to ~/.config/micromanager/config.json, binding a pad key to keys pressed in the focused agent's pane:

{ "chords": { "12": "enter" } }

Why

A text macro composes a prompt but deliberately leaves it unsubmitted — "you still read it and press enter", as the README puts it. That's the right default, but it means finishing a dictated line still requires the keyboard. A chord closes the loop: dictate into an agent pane, then tap the key to submit.

How

HerdrClient.sendKeys already existed and is already used by TuneController to drive the model menu (["up"], ["enter"], ["esc"]), so this is dispatch plumbing rather than new capability. Chord names are passed to pane.send_keys untouched, so the vocabulary is Herdr's — "enter", "esc", "ctrl+alt+u" — and an array is a sequence pressed in order.

A chord outranks a text macro on the same key. The reasoning: the KeyBindings.defaults entries are inherited rather than chosen, so an explicitly configured chord shouldn't have to unbind one first.

Scope is the flexible keys — macroKeyIDs + voiceKeyIDs (9, 10, 11, 12) — the same set that already accepts text macros, addressed the same way including "10+11". The key keeps its KV_OAI_AG* binding, so it stays lit and keeps reporting presses, and nothing touches the device keymap.

Limitations, stated up front

• Chords reach Herdr panes only. They're injected into the pane's pty and never become system-wide key events, so a chord can't drive a system-wide app. Host-side CGEvent synthesis (as VoiceController does for right command) would be the route if that's ever wanted.
• An unrecognised chord name fails at Herdr rather than at config-load time, surfacing through lastError in the panel.

Testing

• 7 new unit tests in KeyBindingsTests: string form, array form, "" unbind, "10+11" expansion, chord-beats-text precedence, and malformed input falling back to defaults.
• Full suite: 89 tests, 3 skipped, and one failure that is pre-existing and unrelated to this change — LiveDeviceTests.testDeviceKeymapHasTheStackKeyBound. It fails on my Creator Micro 2 because the joystick reports type: VENDOR with an empty sectors array, so isAgentKeymapApplied can never return true. Happy to open that separately if it's news; it also means apply() does a flash write on every connect for such a pad.
• Verified end to end on hardware: key 12 lights and submits the focused agent's prompt.

🤖 Generated with Claude Code