docs(log 022): one stem did it

Two complaints weeks apart that sounded unrelated — "the kick doesn't hit" and
"the WAP bass is saturated" — turned out to be the same orbit. d4 is 76-91% of
the low-frequency bed floor on every buried track AND the exposed bass in WAP.
One stem, two jobs, both done badly at different moments, which is what
livecoding does to an orbit and what a set-wide mix decision cannot see.

Punch median 8.3 -> 10.3 dB, buried tracks two -> one, both masters landing on
4786.370000 s so the fifteen approved boundaries survived untouched.

The shareable part is the near-miss: the first render improved every track,
INCLUDING the seven with no correction applied. Good numbers, would have
shipped. What stopped it was asking why the untouched tracks moved — the
balance was attenuating one side instead of holding constant power, so nine
pans were also nine undeclared level cuts being credited to the deliberate
work. After the fix the median went 10.4 -> 10.3: the result was real and the
explanation had been wrong, which is the failure that survives review because
nobody audits a number they like.

armada/tasks/022-one-stem-did-it.md

+---
+log: 022
+title: "One stem did it"
+date: 2026-08-14
+task: "#151 OPAL-26 final mix (also #152, #149, and a samples watcher)"
+tags: [mixing, mastering, opal2026, soundcloud, tooling, salvage]
+shareable: true
+---
+
+## Cap (what & why)
+
+Two complaints, weeks apart, that sounded unrelated. *"The kick doesn't hit as
+loud as rest of set."* And *"the WAP bass is saturated at ~13:10, it's very bad
+at 13:55-14:15."* One is a drum problem, one is a bass problem, and the fix for
+each looked like it lived somewhere else entirely.
+
+They were the same stem.
+
+## Manœuvre (how)
+
+Measured the bed before touching it. `punch` is kick p95 minus low-bed p40, so
+the question "why is the kick buried" reduces to "who owns the bed floor" —
+which stems are sounding at the frames that SET that 40th percentile.
+
+Then built a band-power simulator so a proposed op could be tested in seconds
+instead of an eighty-minute render, and validated the simulator against the
+real master before trusting a single number it produced.
+
+## Prise (findings / artifacts)
+
+- **d4 is 76–91% of the low-frequency bed floor on every buried track.** It is
+  also the exposed bass in WAP. One orbit, two complaints, two different jobs
+  done badly at two different moments — which is exactly what livecoding does
+  to a stem, and exactly what a set-wide mix decision cannot see.
+- Punch median **8.3 → 10.3 dB**, spread 5.3–20.6 → 8.1–23.8, buried tracks
+  two → one. *Am i Doing it Right*, his actual complaint, went 6.9 → 10.7.
+- **Kick-triggered ducking cannot fix punch, and the sweep proved it before it
+  shipped.** p40 lives BETWEEN the kicks, precisely where a kick-triggered duck
+  does nothing: 2–6 dB across three thresholds moved the worst track not at all
+  and made the spread slightly *wider*. A continuous bed wants a static trim.
+- The simulator: r=0.943 against the real master, 1.63 dB residual spread. Good
+  for relative changes, useless for absolute ones, and it says so.
+- Both masters rendered to **4786.370000 s** — the shipped master's length to
+  the millisecond, which is the proof the timeline reconstruction was right and
+  the reason the fifteen already-approved boundaries survived untouched.
+- Along the way, the SoundCloud shelf got audited: of 110 tracks, **only 39 are
+  recoverable as masters**. The other 841 minutes come back as transcodes.
+
+## Sel (the shareable learning)
+
+**The first render improved every track — including the seven I had not
+touched. That was the tell, not the triumph.**
+
+The numbers were good. Punch median up two decibels, spread tightened, the two
+buried tracks rescued. It would have shipped. The thing that stopped it was a
+question with no obvious answer: *why did tracks with no correction get better?*
+
+Because the balance was wrong. I had written it the way a mixing desk writes
+it — attenuate one side, leave the other alone — which moves a stem by turning
+it down. Nine gentle pans were also nine undeclared level cuts of up to 1.1 dB,
+outside the bound the plan declared, and every one of them was quietly being
+credited to the deliberate corrections.
+
+Fixed to constant power, re-rendered, and the median moved from 10.4 to 10.3.
+So the improvement had been real all along and the *explanation* had been
+wrong — which is the failure that survives review, because nobody audits a
+number they like.
+
+A result you cannot fully attribute is not yet a result. The tracks you did not
+touch are the control group, and they are worth more than the ones you did.
+
+## Hameçon (hook)
+
+*"Every track got better, including the ones I never touched — and that was the
+bug report."*