docs(log 024): an address is a lease

The evening the ALSA client numbers swapped and both of the morning's fixes
became wrong at once — while the one tap I had written up as broken turned out
to be the only correct one.

Covers the name-resolved surface tap (#155), the preference-list output router
(#41), the scene keys that could never fire and the one that muted d10 (#134),
and the hush that threw a modal dialog over the editor on a dormant REPL.

The learning is the title: every fix I shipped that morning wrote down a port
number, which looked like a fact and was a lease. Also the smaller trap that
cost twenty minutes — with the surface unplugged, the discovery commands list
no hardware at all, so I nearly 'fixed' a tool that was working perfectly.

armada/tasks/024-an-address-is-a-lease.md

+---
+log: 024
+title: "An address is a lease"
+date: 2026-08-14
+task: "#41 #134 #155 #156 #159 — the evening the client numbers swapped"
+tags: [midi, alsa, routing, pulsar, cockpit, debugging]
+shareable: true
+---
+
+## Cap (what & why)
+
+A morning spent fixing four MIDI taps that were all watching the wrong device.
+An evening watching that fix become wrong again, in the time it takes to unplug
+a USB cable and plug it back in.
+
+The rig is a laptop, a Launch Control XL, an Arturia KeyStep and a Behringer
+interface. In the morning ALSA had the LCXL as client 20 and the KeyStep as 24,
+and three of the four tools tapping MIDI were bound to 24 — reading a keyboard
+nobody was touching, while the surface went unwatched. I re-pointed them all to
+`20:0` and moved on, pleased.
+
+## Manœuvre (how)
+
+That evening PLN replugged, and the kernel handed out the numbers in the other
+order:
+
+    client 20: 'Arturia KeyStep 32'    <- was the LCXL
+    client 24: 'Launch Control XL'     <- was the KeyStep
+
+Both of my fixes were now wrong. The one tap I had *not* fixed — the one I had
+described in a task as broken — was now the only correct one.
+
+So the tap got rewritten to resolve `"Launch Control XL"` **by name, on every
+respawn**, and to stay up whether or not anyone is subscribed, and to keep the
+last value of every control so a view that connects mid-set paints the real
+board instead of a blank one. First run against the live rig, no configuration
+at all:
+
+    starting   port=None
+    ok         port=24:0
+
+## Prise (findings / artifacts)
+
+- **`tools/bridge/surface.py`** — one supervised tap, name-resolved, always on.
+  It also does not litter: three smoke runs of the file each orphaned an
+  `aseqdump` to pid 1, which is precisely the mess PLN had complained about
+  that morning. `try/finally` does not survive SIGKILL, OOM or `timeout`;
+  `PR_SET_PDEATHSIG` moves the guarantee into the kernel.
+- **`tools/master-out.py`** — Ardour's Master follows a preference list
+  (UMC → headphones → speakers) instead of being re-patched by hand in
+  qjackctl on every move. Idempotent by construction: it diffs wanted links
+  against existing ones and applies only the delta, so "already correct" costs
+  zero `pw-link` calls and cannot glitch a live set.
+- **The scene keys never fired, and one of them muted an orbit.**
+  `ctrl-alt-[`/`]` are Pulsar core's fold/unfold, bound on `atom-text-editor` —
+  a more specific selector than the package's `atom-workspace`, so core wins
+  whenever focus is in the editor, which while livecoding is always.
+  `ctrl-alt-0` was `tidalcycles:toggle-mute-10`: "backdrop off" silently muted
+  d10.
+- **A hush on a dormant REPL threw a modal dialog over the editor.** `eval()`
+  boots GHCi first; `hush()` does not, and `remote.js` asserts in a comment that
+  hush is "safe to fire blind". After every Window Reload the first panic-hush
+  was the one guaranteed to fail.
+
+## Sel (the shareable learning)
+
+**An address is a lease, not an identity.**
+
+Every fix I shipped in the morning wrote down a port number. `20:0`. It looked
+like a fact — it had been verified, it was in the task description, three tools
+agreed on it. It was a lease with no expiry printed on it, and the kernel
+renewed it to a different tenant the next time a cable moved.
+
+The tell was available and I ignored it: the thing that had gone wrong in the
+morning *was a stale port number*. I fixed four instances of the symptom and
+none of the cause, then wrote it up as done. It took the numbers swapping —
+the same bug, in reverse, within eight hours — to make the shape visible.
+
+There is a smaller trap folded inside it, which cost twenty minutes on its own.
+While the surface was unplugged, `aseqdump -l` and `aconnect -i` listed no
+hardware inputs at all. I concluded the discovery command was broken and nearly
+"fixed" a tool that was working perfectly. Meanwhile the old taps were still
+running happily on connections made back when the address was valid — so "other
+tools are still reading it" proves nothing about whether the port is right.
+Absence of the device and absence of the listing are indistinguishable unless
+you check whether the device is there.
+
+The general form, for anything that binds to hardware: resolve by the name a
+human would use, re-resolve every time you reconnect, and never persist the
+number. The number is what the system told you once. The name is what you meant.
+
+## Hameçon (hook)
+
+*"I fixed it by writing down the address. The address was a lease, and the
+kernel renewed it to somebody else."*