feat(sokol): Linux evdev/udev + wasm Gamepad API detection sources (Phase 1)

[apotema]

Phase 1 of controller-support epic: labelle-toolkit/labelle-engine#609

Goal

sokol gamepad sources for desktop Linux (evdev + libudev) and wasm (HTML5 Gamepad API), so a sokol build has controller detection — and on Linux, better reconnection than raylib.

Context

sokol has no gamepad layer on any platform. Unlike Android, on desktop Linux an app can open /dev/input/event* directly (subject to permissions), and on the web the Gamepad API is pollable. These plug into the Phase 0 comptime fallback (only compiled in when the backend declares no pollGamepadEvents).

Scope

Linux (evdev + libudev):

1. libudev monitor for add/remove → GamepadEvents into the ring buffer; enumerate existing /dev/input/event* at startup.
2. Derive a stable guid from the device input_id (vendor/product/version + phys) — gives correct reconnection (raylib can't).
3. Permissions: on EACCES, surface via describeGamepads unavailable_reason + a one-time log hint (input group / udev rule). Ship a sample 99-labelle-gamepads.rules (uaccess) and document Flatpak --device=input.
4. Filter duplicate js*/event* nodes and non-gamepad devices (accelerometers etc.).

wasm (HTML5 Gamepad API):
5. gamepadconnected/gamepaddisconnected events → ring buffer; gamepad.id → name, index → slot.

Acceptance criteria

• sokol+Linux: plug/unplug emits events; replug restores same input_id guid; permission failure produces a clear diagnostic, not silence.
• Sample udev rule shipped + documented.
• sokol+wasm: connect/disconnect events propagate.

Non-goals

• Semantic button/axis mapping (the SDL_GameControllerDB problem) is Phase 3. This issue is detection + identity.

[apotema]

Linux runtime data point for the "consolidate desktop Linux onto SDL vs keep udev" question (core#28 follow-up 2, which overlaps this issue's desktop-Linux scope):

The shared backends/sdl_gamepad source is now runtime-verified on Linux — 23/23 checks pass on Ubuntu 26.04 / SDL 2.32.10 (under WSL2, no USB passthrough involved): connect detection through Source.isAvailable, all 15 buttons with correct canonical mapping, all 6 axes including sign conventions, and the synthesized analog-trigger button. Same suite passes on Windows 11 (SDL 2.30.11 MinGW).

This comes from the in-process SDL2 virtual-gamecontroller harness now PR'd in #279 (backends/sdl_gamepad/sim plus a headless CI runtime step), so the Linux behavior is locked in on every PR going forward.

Scope caveat for this issue: the virtual device lives entirely inside SDL's joystick subsystem, so this verifies the source's event pumping / mapping / state logic on Linux SDL — it does not exercise the kernel device path: evdev/hidraw hotplug, /dev/input permissions, or GUID stability across replug. Those are exactly this issue's acceptance criteria. If desktop Linux consolidates onto the SDL source, the parts of this issue that stay uniquely valuable are (a) the wasm Gamepad API source, and (b) the permissions/diagnostics UX — SDL hits the same EACCES on /dev/input and would benefit from the same one-time hint + shipped udev rule.