[bdbarnett]

The pump moves the audio in C, so main() no longer loops printing counters: it starts the card (and its spectrum, on a board with a display) and returns, and the REPL stays usable while it plays. Beside it: stats() prints the pump's counters once, watch() prints them every second until Ctrl-C, and stop() stops the spectrum, the pump and then the amp.

Two things this exposed, both handled:

• A soft reset leaves the C pump running (its buffers aren't on the Python heap), still holding I2S, so the next uac_pump_start() failed with OSError: 259. main() now stops a pump left running before it starts one. The spectrum's meter is safe across the reset: its finaliser detaches it from the sound card.
• stop() then main() failed with "another Meter is attached": stop() now stops the meter and releases it. _spectrum() also treats that OSError like its other reasons for no spectrum, so the card plays either way.

Checked on the ESP32-P4 WIFI6 DEV-KIT (KeDei panel) with music playing from Windows: a cold boot, three stop()/main() cycles at the REPL and three soft resets, each with the spectrum back, no errors, and audio moving at about 100 KB/s afterwards. Each restart counts 6-7 pump timeouts once (the host re-attaching the card) and none after.