docs(notes): PLN's ear on the first machine-trimmed take — "cut it proper and fade"

First listen to the auto-trimmed take-94 mix. The verdict on the sum itself is good
("great mix itself its clean") which clears the record-bus routing of suspicion: 12
orbits summed with a flat -16 dB keeps the balance he actually performed.

The correction is the valuable half. He flagged the silence around 18 min as "not
normal, a bad transition — in a real mixing work wed cut it proper and fade or at
least trim most". That silence WAS detected and reported by take-lens, and survived
anyway: SILENT_RUN_S was 20 s and all three surviving gaps measured 17 s, 18 s and
11 s. A threshold chosen so the trimmer could never eat a musical break had become a
threshold that preserves every bad transition instead.

Two rules extracted, both more general than this take: clamp long silences to a short
breath rather than choosing between keeping all of it and removing all of it; and fade
every cut seam, because every removal this tool has made so far was a butt-join.

Also logged: the stretches he heard as "stuck in loops" are precisely the `loop` kind
the classifier already flags and deliberately does not auto-cut.

armada/tide-table/performance_notes.md

@@ -217,3 +217,37 @@ All three measured + fixed (v2 rebuild, see catch `report.md` + `review_v2_bleed
 - Rotation finally promoted to the shipped kit (v1 drums/bass opened −65/−52 dB
   below peak); instruments re-cut from a roformer-instrumental→demucs CASCADE
   (no rap ghosts: source windows carried vocals at −23 dB).
+
+## OPAL practice take 94 — first machine-trimmed mix (2026-07-29, PLN)
+
+Listening to `mix/take94/take94_trimmed.wav` (19.15 min, uniform −16 dB, 119 s auto-trimmed).
+
+> *"mix is total decent. i hear 3 or 4 tracks well. some moments ive been blocked in
+> loops stuck practice as i was doing other things but tbh great mix itself its clean"*
+
+> *"silence around 18 is not normal its a bad transition ahah in a real mixing work wed
+> cut it proper and fade or at least trim most"*
+
+**What this validates:** the sum + uniform gain trim is clean — no routing fault, no
+bleed audible, the balance PLN performed survives a flat −16 dB. The stretches he
+noticed as "stuck in loops" are exactly the `loop` kind take-lens detects and
+deliberately does NOT auto-cut (hands idle + a couple of orbits droning).
+
+**What this corrects — the rule, in his words.** A silence at a TRANSITION is junk, and
+the mix answer is *cut it proper and fade*, not leave it. The 18 s gap he heard was
+detected and reported, and survived anyway because `SILENT_RUN_S` was 20 s: all three
+remaining gaps (17 s, 18 s, 11 s) sat just under the bar. A threshold tuned to never eat
+a musical break had quietly become a threshold that keeps every bad transition.
+
+Two changes follow, and they generalise past this take:
+1. **Clamp, don't binary-cut.** A silence longer than ~8 s is not a musical rest at
+   120-140 BPM (4 bars ≈ 8 s), so trim it down to a short breath rather than choosing
+   between "keep all 18 s" and "remove every trace". Keeps the seam musical, removes the
+   dead air. This is "or at least trim most" implemented literally.
+2. **Fade every seam.** A butt-join between two takes of a live room is a click, and
+   until now every cut this tool made was a butt-join. Short equal-power crossfades at
+   every removal boundary.
+
+**Ear-gate note for the future:** he heard 3-4 tracks "well" out of a ~21 min take — the
+take has NO track timeline (gig-log gained `k:"track"` the same evening), so his count is
+currently the only boundary data that exists for take 94.