master Take103 (AI ENGINEER set) and Take104 (Rose Rouge club mix)

armada/tasks/2026-09-26-take103-104-mastered.md

+---
+log: n/a
+title: "Take103 (AI ENGINEER set) + Take104 (Rose Rouge club mix) exported and mastered"
+date: 2026-09-26
+task: n/a
+tags: [mastering, postprod, tooling]
+shareable: true
+---
+
+Both takes from `Tidal Live` exported straight from the Ardour capture files —
+no Ardour export pass needed: the interchange `%L/%R` mono pairs were joined
+into 12 stereo orbit stems with ffmpeg (`join=inputs=2`), then the
+`tidal-ears master` chain (analyze → gain-trim → HPF/LPF → sum → glue comp →
+two-pass loudnorm) rendered four masters in
+`~/Work/Sound/Prod/AIEngineer2026_master/`:
+
+- `AIEngineer2026_set_streaming.flac` — **-14.0 LUFS** on target, LRA 8.4
+- `AIEngineer2026_set_club.flac` — -11.5 (single loudnorm can't reach -9
+  against the -1 dBTP ceiling; staged chain is the v2 item)
+- `RoseRouge_clubmix_take104_{streaming,club}.flac` — -13.9 / -10.7
+- `tracks_bandcamp/` — 13 per-track FLACs split with segments authored from
+  the gig-log track-switch timeline (second-precision), titled and dated
+  "AI Engineer, 2026-09-24 (live)".
+
+Honest edges: rec joined *piment_bresilien* 3 min in (take is 76 min, set was
+~95); **d8 was never armed at the gig** (the "Arm Tidal 08" fix happened the
+next day), so Take103 has 11 stems; d9 is digitally silent for the whole Rose
+Rouge take. The track switch log in `gig-20260924-005926.jsonl` proved to be a
+better boundary source than silence detection — instant transitions leave no
+gaps to find.
+
+Found and fixed in tidal-ears (`3bafd03`): `split --segments` formatted the
+slug key as `%02d` and crashed; play-order numbering is now the enumerate
+index. 76 tests pass.
+
+**Shareable bit:** the Ardour capture directory is already the export — mono
+L/R pairs + a JSON timeline of what was played, and the master is two ffmpeg
+passes away. The "export from the DAW" step is optional when the DAW never
+stopped recording.