Fix DJM-900NXS2 USB capture: silent recordings and wire-format distortion

[tonnkatonno]

Summary

The DJM-900NXS2 vendor-class isochronous capture path recorded nothing but digital silence, even with correct MIX/REC OUT routing and confirmed-audible audio on the mixer. Once that was fixed, recordings came out quiet and "washing machine"-distorted. Both are now fixed and confirmed working on a real DJM-900NXS2 + OnePlus CPH2653 (Android 16) over several rounds of hardware testing.

1. Silent capture (was nonzero bytes=0 regardless of routing)

A USBPcap capture of Pioneer's own driver actually recording real audio showed it unconditionally sends a UAC1 SET_CUR sampling-frequency control transfer to the capture endpoint right after SET_INTERFACE, without ever probing GET_CUR first. This app's native code did the opposite: it GET_CUR'd first and silently skipped the SET whenever the GET returned a plausible-looking value -- a static, non-informative response, the same failure shape already known from the MIX-route GET on this model. Sending SET_CUR unconditionally, matching the real driver's sequence exactly, unlocks real audio on the endpoint.

Also added, from the same pcap evidence:

• A native OUT-endpoint duplex keepalive (continuous silence on endpoint 0x01), mirroring the DJM-A9's requiresPlaybackTraffic mechanism, since Pioneer's driver keeps that endpoint active throughout capture.
• MIX/REC OUT routed to both the default output and USB9/10 (output 5) up front, confirmed via a pcap of the Setting Utility toggling both.

2. Wire-format correction (was quiet + distorted after audio started flowing)

Once real audio was flowing, recordings were quiet and cyclically distorted. Testing the real captured wire bytes from Pioneer's driver against candidate channel-count/subframe combinations showed the true format is 12 channels at 3-byte (24-bit) subframes, not the originally guessed 10 channels. Independently corroborated: this app's own captured packets are consistently 216 bytes, which divides evenly into 6 frames of 12ch x 3B (36B/frame) but never evenly into the old assumed 10ch x 3B frame (216/30 = 7.2) -- every packet was being sliced at the wrong frame boundary, which explains the modulating distortion as the sample offset drifted and wrapped every few frames.

Also included

• Fixed auto-pick-loudest-pair channel thrashing: with real signal flowing, several output pairs read comparable amplitude, and the decision was previously re-evaluated on every incoming packet against a still-accumulating peak window, flipping the selected pair dozens of times per second and producing audibly gappy output. Decisions now happen once per finished measurement window, with a hysteresis margin against the currently-selected pair.
• Hardened UsbIsoAudioSource::stop() against a redundant interface release call (NXS2's playback OUT endpoint shares the same interface+alt-setting as capture, unlike DJM-A9's separate one), found next to a destroyed-mutex crash inside libusb_close() during teardown.
• Added a per-output-pair channel selector to the recording UI, mirroring Pioneer's Windows Setting Utility.
• Diagnostics reporting for all of the above.
• Safe handling of Android 14+ foreground-service start rejections that surfaced during testing on a real device.

Test plan

• Confirmed via USBPcap capture that the MIX/REC OUT route SET encoding matches Pioneer's Setting Utility byte-for-byte.
• Confirmed via USBPcap capture of a real Windows recording session that the SET_CUR sequence and 12ch/24-bit wire format match this fix.
• Tested on real hardware (DJM-900NXS2 + OnePlus CPH2653, Android 16): recording produces clean, correctly-leveled, non-distorted audio.
• Tested stop/restart cycles for regressions after the teardown hardening.

🤖 Generated with Claude Code

[P2GR]

Will run a few tests with this later today