[SandroMaglione]

What

Controls: merge devices into one input source

export const Actions = Controls.actions(["up", "down", "left", "right", "interact", "pause"])

export const makeControls = (devices: { readonly keyboard: KeyboardInput; readonly gamepad: GamepadInput; readonly touch: TouchInput }) =>
  Actions.merge({ keyboard: devices.keyboard, gamepad: devices.gamepad, touch: Actions.lacking(devices.touch, ["pause"]) })

const input = resources.input.get()
input.actions.interact.pressed     // merged: held = any holds, pressed = any pressed, released = released and none holds
input.devices.gamepad.left         // each device's own snapshot, keyed by the merge names
input.lastUsed                     // "keyboard" | "gamepad" | "touch" | null
input.available.pause              // false while the last used device lacks it
services.controls.rebind.gamepad({ ... })

• Compile-time checks: each of these fails to compile:
  • a device that binds an undeclared action;
  • a device that misses an action without listing it in lacking;
  • a lacking list that names too much or too little (autocomplete offers exactly the missing actions);
  • an action that no device binds;
  • reading a lacking action on that device (input.devices.touch.actions.pause).
• Lacking actions in describe(): each one is reported as a service-lint.

One snapshot shape (breaking)

• Keyboard: the snapshot is now { actions, used }.
• Touch: buttons is renamed to actions and touched to used.
• Gamepad: gains used.
• What counts as used: a press or a hold, or a stick or swipe in use. Releases don't count, so a blur or a disconnect doesn't change lastUsed.

Checked capture

• New access kind: Game.System.consumeService, plus a service-consumed-twice lint that fires when one input is captured in two schedules and each capture sees only part of the presses.
• InputCapture.system: keeps its real spec. A thisTick reader placed before it is now a compile error that names both systems.
• Debug.serviceLints: a hook through which service values report their own lints.

Top-down now reads one merged resource instead of two.

Verification

• pnpm run check passes.
• pnpm pack:check passes.
• pnpm bench:check flags query/iterate-write-10k at +26.5% against the committed baseline. Compared with origin/main on the same machine (--compare), there is no regression, and this PR touches no query code. Type-checker metrics are +2.6% against the baseline; main is already at +2.5%.

🤖 Generated with Claude Code