docs(log 021): the editor knew all along

Captain's log for #148. PLN heard the album's cuts were wrong within seconds of
a private SoundCloud upload; the bug went four layers deep, ending at a
provenance note that borrowed a true r>=0.99 from an entirely different
measurement.

Two learnings worth the write-up. First: a verifier that passes the case you
built it for is not yet a verifier — mine passed WAP and Drums, the two tracks
he had actually complained about. Second, and his line not mine: the text editor
knew every boundary exactly and never wrote it down, so we spent an afternoon
reconstructing from audio what a log line would have given for free.

armada/tasks/021-the-editor-knew-all-along.md

+---
+log: 021
+title: "The editor knew all along"
+date: 2026-08-13
+task: "#148 album cuts wrong (also #149, #150, #151)"
+tags: [mastering, tooling, opal2026, boundaries, salvage]
+shareable: true
+---
+
+## Cap (what & why)
+
+PLN uploaded the OPAL-26 album privately to SoundCloud to test it, and heard the
+problem in about ten seconds: *"cuts are wrong! WAP starts on the last sound of
+prev, then silence, then first WAP sound at 00:07? the next track, something
+about drums, starts on a wap sound??"*
+
+He asked for the bug chased "down to its bottom source and throughout erroneous
+layers". There were four.
+
+## Manœuvre (how)
+
+Checked the arithmetic before the audio. Every one of the fifteen track
+durations in the shipped segments file equalled a gig-log tab-switch delta to
+within 0.01 s — so the boundaries had never touched the audio at all. Then
+measured the log-to-audio anchor honestly, by correlating three fader CC curves
+against their own stems.
+
+Then built the detectors: silence, per-orbit activity, and eventually tempo.
+Then built the thing that was actually missing, a verifier.
+
+## Prise (findings / artifacts)
+
+- **The gig log's `track` event is a TAB-FOCUS event, not a music event.** PLN
+  opens the next track's file while the current one is still playing, reads it,
+  edits it, evaluates it seconds later. The editor knows when he LOOKED at a
+  track; only the audio knows when it STARTED. Lag measured **+1.0 s to +50.7 s,
+  always the same direction**.
+- The anchor was 1.1 s off as well. True offset 745 s, measured three ways, all
+  agreeing.
+- **The layer that hid it:** the provenance note claimed the boundaries were
+  "aligned to audio by cross-correlation, r>=0.99, 20 probes, 0.00 ms spread".
+  Every word true — of a *different* measurement. That r>=0.99 was the
+  take-to-master audio alignment. It never validated the log-to-audio anchor.
+- **Silence is the weakest evidence.** Shipping "a real gap is authoritative"
+  put four boundaries 15-47 s early, because a livecoded track is full of holes
+  and a breakdown inside the outgoing track outvoted the orbit evidence.
+- **Tempo was the lens that worked**, and PLN asked for it as a third detector
+  once he saw it. Not by estimating tempo — that is unstable at exactly the
+  transition you care about, and a 128 BPM track read 171.4 for long stretches —
+  but by scoring the onset flux against both KNOWN BPMs from the setlist and
+  finding the crossover. It caught a handover that gap and orbit-flip both
+  missed by 47 seconds.
+- Verifier scores: shipped boundaries **3/15**, refined **12/15**.
+- Along the way, two taste complaints where the obvious metric said he was
+  wrong: "the kick doesn't hit" (level said +1.3 dB ABOVE median; PUNCH said
+  buried) and "the bass is saturated" (nothing is clipped; everything that was
+  masking it left).
+
+## Sel (the shareable learning)
+
+**My first verifier passed the two tracks PLN had complained about.** I wrote it
+specifically to catch his bug, ran it, got a satisfying 3/15 failure report, and
+only noticed on a second read that WAP and Drums were both in the passing
+column. A verifier that passes the case you built it for is not yet a verifier —
+it is a hypothesis wearing a lab coat.
+
+And the bigger one, which he put better than I did: *"tbh why dont we have
+events when i ctrl+enter a track? the hud changes, we could know when one is in
+focus no?"*
+
+We spent an afternoon reconstructing from audio something the editor knew
+exactly, at the millisecond, and threw away. Every clever detector in this log is
+compensation for a missing log line. The instrument that measures best is the
+one standing closest to the event.
+
+## Hameçon (hook)
+
+*"Four detectors, three afternoons, and a verifier that agreed with the bug —
+to recover a fact the text editor had known all along and never bothered to
+write down."*