[edumeneses]

PipeWire gives no buffer to an output port that nothing is linked to. fetch_cycle_buffers leaves those outputs null, and assign_chunk_pointers pointed every one of them at the same m_scratch. So whatever the tick wrote to one unlinked output landed in all the others.

Nothing reads those ports, so nobody hears it. But the hardware output meters read the tick's outputs after the mix, and they show it. Edu hit this testing the mixer (ossia/score#2302) with score's outputs left unlinked:

• a signal sent to /out/1 showed on every channel's meter;
• a stereo source, with three outputs, showed on the third one too.

With 64 outputs and a stereo default sink, channels 3 to 64 all share one buffer in the same way. The per-channel /out/N gains of #950 also scaled that shared buffer once per alias.

The fix:

• m_scratch holds one block per output.
• assign_chunk_pointers takes a stride, so the i-th missing output writes at scratch + i * stride.
• The inputs keep their single shared m_silence: the stride defaults to 0.

Tests, in score (they live there with the other PipeWire tests). They are in ossia/score#2336, which bumps this:

• test_unit_pipewire_quantum: two missing channels get distinct scratch.
• test_unit_pipewire_protocol_quantum: a private daemon with no sink and three unlinked outputs. Each output must still hold what the tick wrote to it. Without the fix it fails: 18 times, an output held another channel's signal at the end of the tick.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lc7tYGHRn6TmaxMEfBTswW

[jcelerier]

lgtm, thanks!