Parked in debug-log.md (2026-06-10). The I2C interrupt-storm fix (render blocks on new data + dirty-check) resolved the "decoder is too slow" starvation by cutting the interrupt load that was preempting MPD's audio thread.
Belt-and-suspenders follow-up: give MPD's output/decoder thread real-time scheduling priority (or a negative nice) so a future runaway on the Pi can't starve it again. Not needed today — the interrupt reduction alone fixed the observed problem — but cheap insurance if other load lands on the box.
Parked in
debug-log.md(2026-06-10). The I2C interrupt-storm fix (render blocks on new data + dirty-check) resolved the "decoder is too slow" starvation by cutting the interrupt load that was preempting MPD's audio thread.Belt-and-suspenders follow-up: give MPD's output/decoder thread real-time scheduling priority (or a negative nice) so a future runaway on the Pi can't starve it again. Not needed today — the interrupt reduction alone fixed the observed problem — but cheap insurance if other load lands on the box.