USB iso OUT: pace packets by audio rate, not wMaxPacketSize

[patrickrb]

Summary

libusb_set_iso_packet_lengths was being called with the endpoint's wMaxPacketSize (200 bytes for a C-Media CM108 at 48 kHz stereo 16-bit) but USB Audio Class devices play back exactly however many bytes per frame the host hands them. Sending 200 bytes per 1 ms frame at a negotiated 48 kHz rate makes the device clock samples out 200/192 = 4.17 % faster than negotiated — every audio tone gets shifted up by that ratio.

For FT8 specifically:

• 1000 Hz audio offset → 1042 Hz on the air
• WSJT-X tone spacing 6.25 Hz → ~6.51 Hz (off the decoder's tone grid)
• 12.64 s message duration → 12.14 s real time

That last number matches the observed nativeWrite return timing exactly, and the off-grid tones are why receivers couldn't decode our signal even though the rig was happily putting out clean ALC-correct 70 W of properly modulated SSB. The Android-standard AudioTrack path was unaffected because the kernel UAC driver computes the per-frame byte count from the negotiated sample rate.

Fix

Use the three outputSampleRate/outputChannels/outputBytesPerSample parameters (which were always being passed from Java but ignored by nativeWrite) to compute the correct per-frame byte count for USB FS:

int bytesPerFrame = (outputSampleRate * outputChannels * outputBytesPerSample) / 1000;

Apply it to both buffer allocation and libusb_set_iso_packet_lengths. Fall back to maxPacketSize (with a loud log) if the format args are bad, so unrecognized devices degrade to wrong-pitch audio rather than a silent abort.

Test plan

• Pixel 8 + FT-891 + C-Media CM108 (0D8C:0012): nativeWrite for 12.64 s of audio used to return at +12.14 s; now returns at +12.64 s (verified in debug.log timing).
• FT8 tones now land on the WSJT-X grid → spots appear on PSKReporter when transmitting CQ on 14.074 MHz.
• Sanity check: car-dash USB CODEC path (the original libusb iso target from PR USB Audio: libusb iso OUT for TX (companion to #83) #84) still works.

Notes

This is independent of #93 (Costas chop) but both are required for direct-USB-audio TX to actually decode. The author tested the local build with both PRs applied; this PR text only covers the iso fix.

🤖 Generated with Claude Code