Every intuition on the way to this one was wrong, including several of mine, so the
log records the wrong turns as well as the answer — that is the part worth reading
back.

The shape: nothing crashed. sclang up, scsynth up on 6.7% CPU, sink RUNNING,
routing intact, and no sound. Any monitor aimed at SuperCollider would have
reported a healthy rig. PLN found the discriminator himself by switching the
default sink by hand — internal speaker plays, UMC does not — which killed the
pure-CPU theory and pointed at the interface, which was also not at fault: 128
vs 1024 period_size is an 8x tighter deadline, so the UMC is the canary.

Recorded because they will recur:
  - `state: RUNNING` lied while the device was dying once a second; only repeated
    hw_ptr sampling is honest, and FOUR samples, not two — my two-sample read
    called it "still stalling" when a ring-buffer wrap looks identical to a stall.
  - a rule scoped to one device but applied to the whole graph stops being a fix
    and becomes a load (52-no-suspend matched every node; two nodes nobody used
    held 2.1M xruns).
  - the laptop mic had auto-wired itself into the d1 stem, invisible in Ardour's
    GUI and in the saved session, and Take95 is already exported (#132).
  - #7 survived two fixes because the metric counted the BOUNDED map.
  - my own journalctl scans burned a core twice while he was losing audio. I was
    the load I was measuring.