USB Audio: libusb-backed iso capture (fixes car-dash)

[patrickrb]

Summary

Replaces Android's broken UsbRequest isochronous path with a libusb-backed native capture for "(USB direct)" audio input. This is the actual fix for the car-dash Android 11 tablet, which has been falling back to the built-in mic since #82's bail-out kicked in (because UsbRequest.queue() throws IllegalStateException: request is not initialized on its first call there).

Two commits — toolchain plumbing landed first, then the libusb capture on top — so the diff stays reviewable.

Why libusb

Android's UsbRequest API in theory supports isochronous transfers since API 26, but in practice the AOSP wrapper's iso path fails silently on some kernels — notably automotive Android skins. UsbRequest.initialize() returns false, the next queue() throws, and the user never sees real audio. Phase 0's diagnostics (#82) caught this in the wild.

libusb talks to USBDEVFS via raw ioctls; it's the most battle-tested USB stack on Linux and ships with explicit Android support (libusb_wrap_sys_device(fd)).

What's in it

Phase 1 — NDK / CMake toolchain

• Re-enables the previously-commented-out externalNativeBuild { cmake { path } } in app/build.gradle.
• Drops cppFlags '-JENABLE_XOM=false' inherited from upstream FT8CN. It's a typo (likely intended -D, not -J); clang rejects it once a real native build is wired up. Was a silent noop before because libft8cn.so is shipped prebuilt via jniLibs.srcDirs.
• Adds src/main/cpp/CMakeLists.txt building a new SHARED library libft8af_usb.so — distinct name from libft8cn.so so the prebuilt FT8 decoder coexists unchanged.
• UsbAudioNative.java: System.loadLibrary wrapper, isAvailable() guard. Load failure logs and returns false rather than crashing.
• ComposeMainActivity.onCreate logs the native sentinel right after === APP START === for one-line confirmation that the JNI bridge is live.
• ndk { abiFilters } locked to the four ABIs we already ship.

Phase 2 — libusb capture

• Vendors libusb 1.0.27 sources under cpp/libusb/. Non-Android subdirectories (Xcode/, msvc/, doc/, examples/, tests/, CI configs) removed up front to keep the repo lean. What's left is what android/jni/libusb.mk compiles plus the upstream Android config.h. License is LGPL-2.1 (compatible with our MIT app license; upstream credit retained in cpp/libusb/AUTHORS, COPYING, README).
• CMake builds libusb as a static lib (10 .c files mirroring android/jni/libusb.mk) and links into libft8af_usb.so.
• usb_audio_capture.cpp: JNI nativeStart / nativeStop. Java passes the fd from UsbDeviceConnection.getFileDescriptor() plus endpoint/interface params; native wraps with libusb_wrap_sys_device(), allocates 4 iso transfers × 8 packets each, and pumps a libusb_handle_events_timeout_completed loop on a worker thread. Each completed packet's int16 PCM is converted to mono float and decimated to the target rate before calling back into Java.
• UsbAudioDevice.startCapture() prefers the libusb path. On success, the original captureLoop() is bypassed; on failure (handle == 0) it falls back to the UsbRequest loop so devices that were already working don't regress.
• stopCapture() routes through nativeStop when a native handle is live — blocks until the event thread has cancelled all outstanding URBs.
• Every transition logs to debug.log: which path was tried, the parameters, and whether it started/stopped cleanly. Easy to read in the in-app Debug screen.

Test plan

• Phase 1: Pixel 8, build green, libft8af_usb.so packaged for all four ABIs, sentinel line in debug.log after install.
• Phase 2: build green, libusb compiles for all four ABIs from a clean checkout (no errors, no warnings beyond the existing Kotlin unused-var ones). APK contains libft8af_usb.so 298KB–516KB depending on arch (libusb statically linked).
• Pixel 8 smoke test after merge: confirm (USB direct) selection still captures audio normally — the libusb path should succeed, not fall back to UsbRequest. Look for UsbAudioDevice: libusb capture started OK in debug.log.
• Car-dash Android 11 tablet (after Play update): pick (USB direct), run waterfall. Expected outcome: waterfall fills with real RF instead of cabin mic. Diagnostic line in debug.log should read UsbAudioDevice: libusb capture started OK rather than the IllegalStateException + giving up on raw USB pair from Fix USB audio death-loop; show real version in splash & About #82.

If libusb itself can't drive iso on the car-dash

We'll see one of these in debug.log and then know we need a different fix (e.g., try a non-iso alt-setting if the device has one, or a USBDEVFS_SUBMITURB-direct approach):

• libusb_wrap_sys_device(fd=...): LIBUSB_ERROR_*
• submit transfer i=0: LIBUSB_ERROR_*
• iso transfer status=... (non-fatal; resubmitting) repeating forever

🤖 Generated with Claude Code