gig: the set that logged itself, and the 17m44s before REC

Reconstruct 2026-09-24 from four independent lenses: gig-log eval, the
lcxl3-driver journal, gig-log focus, and full-resolution per-second levels
off the eleven Ardour stems. Eval and the surface journal agree to the
second on all fifteen boundaries, so nothing here is a statistical guess
the way CosmicFest's recovery had to be.

What the numbers say: 94 minutes, all fourteen planned tracks in order plus
sunny_side_up sixth, five xruns, zero throttle events. REC landed 17m44s
into the set, so salut_nu and love_first have no audio at all. Orbit 08 was
never record-armed and the control map lists it in every track. Neither
MIDI recorder captured a single control move.

armada/tasks/048-the-set-that-logged-itself.md

+---
+log: 048
+title: "The set that logged itself"
+date: 2026-09-25
+task: "n/a — post-gig reconstruction, 2026-09-24"
+tags: [tooling, salvage, mastering, gig]
+shareable: true
+---
+
+## Cap (what & why)
+CosmicFest-26 had to be recovered from one phone recording by three statistical
+lenses, and six of eighteen rows came back `unresolved`. Thirteen months of
+instrumentation later, Thu 24 asked the same question — *what did I actually
+play, and what got recorded* — and this time the rig already knew.
+
+## Manœuvre (how)
+Four lenses, none trusted alone. gig-log's `eval` (a real ctrl+enter reached
+GHCi) and the `lcxl3-driver` journal's `track ->` lines agreed **to the second**
+on all 15 boundaries, so the timeline needed no fusion verdict at all — two
+independent witnesses saying the same thing. The Ardour stems were then only
+*measured* inside those windows, never used to guess a name.
+
+One trap on the way: the first level pass read 4096 of every 48000 frames and
+reported most orbits at 0% active. That was the window landing between hits, not
+silence. Reading every sample of a 19 GB take costs 28 seconds. Decimation on
+percussive material is not an optimisation, it is a wrong answer.
+
+## Prise (findings / artifacts)
+- `armada/tide-table/build_gig_reconstruction.py` — re-runnable, four lenses in.
+- `take103_2026-09-24_recovered_set.md` + `.json` — the set, per-orbit levels.
+- **Set: 18:09:57 → 19:44:06, 94 min, 15 tracks.** All 14 planned, in order,
+  plus `sunny_side_up` inserted 6th — the encore candidate the setlist had
+  deliberately parked.
+- **5 xruns in 94 minutes. 0 throttle events.** The session's other 1177 belong
+  to soundcheck and the two laptop-sink nodes nobody cut. Four earlier sessions
+  logged 5172 / 4852 / 1114 / 987 Ardour xruns; the thread demotion holds.
+- **REC at 18:27:41 — 17m44s late.** `salut_nu` and `love_first` exist in no
+  audio; `piment_bresilien` starts 3:10 in. 81% of the set on disk.
+- **Orbit 08 was never armed** (`rec-enable value="0"`, alone among 12), and the
+  control map lists it in all 15 tracks. No file was created — an absent stem,
+  not a silent one.
+- **No MIDI from either recorder.** gig-log bound seq port `133:0` at 00:59 and
+  never rebound; Ardour's `Keys` wrote 24 files of 26 bytes each.
+
+## Sel (the shareable learning)
+**The lens that saved the night was not built for it.** gig-log has a MIDI lens,
+a recording lens, an eval lens, a thermal lens — and the reconstruction stands on
+the *LED driver's log lines*, written to explain surface repaints. Meanwhile the
+purpose-built MIDI lens had been dead since the night before and the purpose-built
+OSC recording lens was wrong for the entire 76 minutes (`conflict: true` from the
+first record to the last: Ardour's own transport feedback never once said
+"recording" while it was recording).
+
+Redundancy is not three copies of the good instrument. It is four instruments
+that fail differently.
+
+Corollary, learned the same morning: **the session file no longer holds the mix
+you played.** Nine of twelve faders read −∞ in `Tidal Live.ardour`, saved at
+19:44:41, twenty-six seconds after a surface `lap reset` zeroed the board. The
+stems prove those faders were up — orbit 04 sits at a −17 dBFS median and is the
+loudest orbit in 13 of 15 tracks. Audit the artifact, not the config.
+
+## Hameçon (hook)
+"I forgot to hit record." Ninety-four minutes later the laptop could still name
+every track he played, to the second, from a log written by the thing that paints
+the LEDs.
+
+## Sillage (what it unlocks)
+- Mastering can start: 11 stems, 76m19s, t0 = 18:27:41.201, per-orbit levels done.
+- Peaks over 0 dBFS on 13 of 15 tracks (worst +7.7) — the limiter was working.
+  His ear decides whether that is the sound.
+- A REC reminder is now buildable **and** the lens it must not use is identified.
+- Two open holes with names: orbit 08's arm, and gig-log's MIDI rebind.

armada/tide-table/performance_notes.md

@@ -743,3 +743,38 @@ no number. The launcher now reads the knobs and names their owner.
 a thirteen-minute silence caused by a diagnostic SuperDirt restart. Worth
 recording as pacing data: the gig-eve session that produced the most repairs also
 produced the least playing, and he called it himself rather than grinding.
+
+---
+
+## 2026-09-24 — the AI ENGINEER set (Take103)
+
+**"boom, was a good live!"** and **"overall great gig bro"** — the whole-set
+verdict, unprompted, the morning after. First unqualified positive on a full
+two-hour slot. Worth holding against the measurements: 5 xruns in 94 minutes,
+zero throttle events, all 14 planned tracks in the planned order. The night the
+rig stopped being the subject.
+
+**"the midiviz had great success"** — audience reaction to the MIDI lens, not his
+own. Third independent signal that the *visual is part of the instrument*,
+after "the midi mon is beautiful" and the `salut_nu`-has-no-backdrop complaint.
+The taste is consistent and it is not about his comfort: he wants the visual
+layer treated as a deliverable of the performance, because the room responds to
+it. Note the irony worth telling: the surface the audience loved is the one whose
+own log captured nothing — gig-log's `cc` stream is empty for the entire set.
+
+**"i had forgotten to REC at first (we can discuss how to make that reminded!)"**
+— measured: REC landed at 18:27:41, the set started 18:09:57, so **17m44s** and
+two whole tracks (`salut_nu`, `love_first`) exist in no recording. He raised it
+himself before being told, and asked for a reminder rather than a fix. Pacing
+signal: the miss is at the *start* of the set, where attention is on the room and
+the first pattern, not on the DAW.
+
+**He played `sunny_side_up` 6th, off the setlist.** `setlist_thu24.txt`
+deliberately parked it as the encore candidate ("the most-played thing in the
+catalog, and it is NOT in the 14"). It went in at 18:42:41, right after
+`bain_electrique`, the set's fastest. Not a mistake to log — a live decision, and
+the only deviation from the written order in 94 minutes. Ask him what it was for:
+a lift after the 128 BPM, or a reach for something safe?
+
+**`rose_rouge` ran 11m34s**, nearly double the set's median track. The closer
+took its time. Consistent with the "closes on the slowest" design note.