docs(log): 019 — the ghost that cut the new track dead, and 32 buttons only an ear could catch

Captain's-log entry for the 2026-07-29 session, written while it is fresh (the
armada/tasks discipline: these logs are the raw material for posts and video
narration, not bureaucracy).

Two findings, one class: a machine check stayed green while the instrument was
broken. One was Tidal's (xfadeIn turns the outgoing pattern down but never off, and
a gain-0 event still broadcasts its cut group). One was mine, shipped the previous
evening (the button remap assigned cells in reading order, inverting 32 gate/gesture
pairs that only PLN's ear caught).

The shareable core, for later: a gain of zero is not the absence of an event. In a
system where events carry side effects — voice stealing, global FX bus writes — the
silent ones still act. Plus the green-check problem twice in one day, convergence as
a bug report, and the honest limit of verdict-equality checking.

armada/tasks/019-the-ghost-that-cut-the-new-track-dead.md

+---
+log: 019
+title: "The ghost that cut the new track dead — and 32 buttons only an ear could catch"
+date: 2026-07-29
+task: "#102 #94 #100 (with #13, #35, #84, #99)"
+tags: [tooling, gig, opal2026, measurement, livecoding, tidal-internals]
+shareable: true
+---
+
+## Cap (what & why)
+
+OPAL is J-7 and PLN was recording takes, so every finding arrived as a symptom
+heard mid-set rather than a bug report. Two of them turned out to be the same
+class: **a machine check had been green while the instrument was broken.** One was
+Tidal's fault. One was mine, shipped the previous evening.
+
+## Manœuvre (how)
+
+**The fade.** PLN had been losing orbits all afternoon — *"when i leave d5 on the
+33 effcet, it fades in oblivion!"* — and had found a workaround before I found the
+cause: *"so for now hush then ctrl+enter seems like a fix ahah but not a lovely one
+:D"*. The sentence that cracked it was his: *"reboot-then-ctrl-enter: doesnt seem
+to obliviate, loops forever. sounds like a track-to-track state bug"*. Same code,
+two outcomes, differing only by **pattern history** — and `streamHush` prepends
+`silence` to exactly that history.
+
+From there it was reading, not guessing. Tidal's `xfadeIn` is one line; `|*` takes
+structure from the left; so the outgoing pattern keeps every event it ever had.
+Confirmed by `queryArc` rather than by argument, then cross-checked against
+SuperDirt's `DirtEvent.sc` to find out whether a gain-0 event is actually harmless.
+It is not.
+
+**The buttons.** PLN heard it twice in ten minutes: *"techno drum mask is inverted
+on d8"*, then *"drums in revolution seem way too fast ... yea def an inversion"*.
+I measured the damage on my own commit's diff instead of on the corpus, which is
+the difference between "how many odd sites exist in 826 files of history" and "how
+many did I just break".
+
+## Prise (findings / artifacts)
+
+**The ghost (`be52627`).** Every `dN` is a 4-cycle `xfadeIn`. It turns the outgoing
+pattern DOWN but never OFF: measured, **at cycle 64 of a 4-cycle fade it still
+emits one onset per cycle, at gain 0**. And SuperDirt broadcasts the cut group in
+`playSynths` (`DirtEvent.sc:182-189`) *before* amplitude is ever read — `~amp`
+appears later, at line 160, as a gate-synth argument — with `cutAllSamples: 1` for
+a positive `cut`. So the silent ghost kept releasing the incoming pattern's voice
+on every single hit. The orbit was not fading out. **It was being cut to death by
+its own predecessor.**
+
+Onsets per cycle, incoming/outgoing, fading at cycle 0 over 4:
+
+```
+xfadeIn 4    1/1 1/1 1/1 1/1 | 1/1 1/1 1/1 1/1   ghost forever
+clutchIn 4   0/1 0/1 1/0 1/0 | 1/0 1/0 1/0 1/0   clean, never overlaps
+xfadeCut 4   1/1 1/1 1/1 1/1 | 1/0 1/0 1/0 1/0   clean from cycle 4
+```
+
+Fix = `playFor now (now+t)` on the outgoing pattern, envelopes untouched.
+`clutchIn` is arguably better for a cut-heavy rig; it is deliberately NOT the
+default, because changing the feel of every transition 7 days before a gig is not
+a repair.
+
+**The buttons (`5089eec`, `2c178c9`, `6ed637d`).** Phase 2 of the remap assigned
+each orbit's two button cells in *first-appearance order*. Tracks write the gesture
+line above the gate line. So: **32 inversions, 13 files**, nearly all in the
+setlist. 81 cells then moved onto the row their role belongs to, across 39 files,
+with `silent-eval --seeded` verdicts byte-identical twice and alignment still
+83/83.
+
+**Two lints PLN asked for by name.** PV010 (role on the wrong row) and PV011 (one
+FX bus slot shared by two orbits — *"like shared busses they screw things"*). PV011
+immediately found `crushbus 41` shared by take_5's bass and choir.
+
+**Delegated, landed:** the per-stream highlight colours (`2c0524e`) and the HUD
+telling the rig which track is loaded (`2576e77`).
+
+## Sel (the shareable learning)
+
+**A gain of zero is not an absence of an event.** That is the whole bug. Turning a
+pattern down is not the same as stopping it, and in a system where events carry
+side effects — voice-stealing, global FX bus writes — the silent ones still act.
+Tidal's crossfade is correct about loudness and wrong about consequence.
+
+**The green-check problem, twice in one day.** `check-boot.sh` had four passes and
+all four stayed green on the broken transition: it typechecked, it parsed the way
+Pulsar chunks it, and it emitted events. The only check that can see this bug is
+one that *counts onsets from the outgoing pattern after the fade* — so that became
+Pass 5, and it is mutation-tested, because a gate you have never seen fail is not a
+gate. Meanwhile the highlight wiggle had been "fixed" weeks ago in a stylesheet
+that could not finish the job: the user copy won the cascade per-property, but **a
+cascade winner cannot unset a property it never mentions**, so the `border` in the
+package sheet underneath kept reflowing the line.
+
+**Convergence as a bug report.** The role classifier was wrong three times, and
+each time the tell was the same: apply the repair twice and it disagreed with its
+own previous pass. `mask . chop` is a gate. `superimpose (struct ...)` is NOT — the
+`struct` is two levels down building the added layer. And a body that opens on the
+next line reads as "unknown", which is dangerous rather than neutral, because an
+unknown control gets a positional fallback that re-creates the inversion.
+
+**And the honest limit of the gate I trusted.** All 81 moves passed verdict
+equality — soundly, but narrowly. Every button CC seeds to 0, so permuting CCs
+between two calls in one block cannot change either default branch *by
+construction*. When a swap turned `("^59" - "^91")` into `("^91" - "^59")` I nearly
+"fixed" a non-bug; the check could not have told me either way, because `0-0=0`
+whichever order it is in. **Verdict equality at the seed answers "did this silence
+anything at boot" and nothing else.**
+
+## Hameçon (hook)
+
+"My livecoding rig kept killing tracks the moment I loaded them. The fix was `hush`
+first — and the reason was that the *previous* track was still playing, silently,
+murdering the new one on every beat. A gain of zero is not silence. It's an assassin
+with the volume down."
+
+## Sillage (what it unlocks)
+
+- #102: PLN reboots and records a full set — the one test only ears can run.
+- #13 (transitions) has a solid floor now, so the remaining work is craft: choose
+  per-situation transitions, and decide post-OPAL whether `clutchIn` should be the
+  default for a rig where most orbits own a cut group.
+- #77 is untouched by all of this and now clearly separate: an orbit the incoming
+  track never *declares* gets no eval, so no transition, so it just keeps playing.
+- take_5's d4 fade is still unexplained — it has no cut group, so at least one more
+  mechanism exists. Named rather than absorbed into the win.