[karstx]

Adds hidra::webusb, a wasm backend built on nusb's WebUSB platform, for devices the browser will not expose over WebHID.

Stacked on #8.

Why

WebHID only surfaces interfaces the host recognises as HID. A device whose interface declares a vendor-specific class never appears there, even though it typically still speaks the HID protocol on the wire.

The split is forced by the platform: Blink refuses claimInterface on the protected classes — HID (0x03), Audio, Video, Mass Storage, Smart Card, Wireless Controller. So the two backends are complementary rather than interchangeable:

interface class	WebHID	WebUSB
HID (0x03)	yes	no — SecurityError
vendor-specific (0xFF)	no — invisible	yes

This is why the backend deliberately does not filter for HID-class interfaces the way the native nusb backend does: those are precisely the ones it cannot have, and WebHID already covers them.

What

• src/backend/webusb.rs — the HID-over-USB mapping, mirroring the native nusb backend: SET_REPORT/GET_REPORT class requests for feature and input reports, interrupt OUT for writes (falling back to SET_REPORT(Output)), interrupt IN for reads, GET_DESCRIPTOR(Report) where the interface has one.
• hidra::webusb::{Hidra, Device, HidDevice, DeviceSelector} — async throughout, since the browser event loop cannot block. request_device drives the chooser; open claims an interface (None takes the first).
• Endpoints sit behind a take/put RefCell: wasm is single-threaded but futures still interleave, so a second in-flight transfer is refused rather than aliasing the endpoint.
• Gated on feature = "nusb" + wasm32, so the WebHID-only build is unchanged.

Incidental

The crate-level docs linked native-only items (Backend, Hidra::builder) unconditionally, so every wasm doc build failed. Pre-existing, and invisible because the docs job only documents the native target. The platform paragraphs are now cfg_attr-gated, and the wasm CI job additionally checks the --no-default-features configuration.

Testing

fmt, clippy (native, wasm default, wasm no-default), test, MSRV 1.85 and both doc targets pass locally.

No e2e leg yet: exercising WebUSB needs either a real vendor-class device or browser-side emulation, so it is not covered by the virtual-device suites. Worth adding once there is a device to point it at.

[karstx]

Closing: the devices this was built for cannot be reached over WebUSB.

Tested against a real plugged-in Sinowealth mouse (258a:0036) in Chrome with WebUsbAllowDevicesForUrls granting it. The backend itself behaves correctly — it enumerated the device by name through device_list and reported a clear error — but the claim is refused by the browser:

An attempt to claim a USB device interface has been blocked
because it implements a protected interface class.

Every archived Sinowealth ISP bootloader declares bInterfaceClass 0x03 (0904 0000 0003 0000 00, bNumEndpoints 0), which is precisely the class Blink blocks. WebUSB is only a way in for devices declaring a non-HID interface class; it cannot substitute for WebHID when a HID-class device is unreachable, which was the premise here.

The branch stays if a genuinely vendor-class device ever turns up. What it contained:

• hidra::webusb, mapping reports onto control and interrupt transfers via nusb's WebUSB platform
• a webusb e2e target: a dummy_hcd + configfs FunctionFS gadget declaring class 0xFF, driven through Chromium/Playwright, covering three device shapes (full, control-only, no-report-descriptor) — all passing in the Linux test VM

One fix here is worth salvaging and is unrelated to WebUSB: the crate-level docs link native-only items (Backend, Hidra::builder) unconditionally, so every cargo doc --target wasm32-unknown-unknown build fails. That is pre-existing on master and invisible in CI, because the docs job only documents the native target. Happy to land the cfg_attr gating as its own PR.