docs(log 020): the master that was already clipped

Captain's-log entry for the OPAL-26 mastering session (#146). The documentary
beat: three separate times today something SOUNDED fine while its metadata lied
— 11 of 12 shipped stems hard-clipped, every master this tool ever made silently
192 kHz, and fifteen split files each claiming to be the whole 79-minute set.
None of it errored. An audio pipeline does not fail by crashing; it fails by
producing a file that opens.

Counterweighted with the hour I spent building an elegant theory about a
"lingering arp" around an orbit measuring -72.6 dBFS, which PLN closed in forty
seconds by listening.

armada/tasks/020-the-master-that-was-already-clipped.md

+---
+log: 020
+title: "The master that was already clipped"
+date: 2026-08-12
+task: "#146 OPAL-26 mastering (also #91, #144, #145)"
+tags: [mastering, tooling, opal2026, salvage]
+shareable: true
+---
+
+## Cap (what & why)
+
+ParVagues played Opal Festival 2026 — "Sunset Forest", Scène Cabane, sunset
+slot, 15 tracks in 1h20. PLN did a quick Audacity master and asked for two
+things: a critique of the set track-by-track *before* he shared his own notes,
+and then a real master. Plus metadata, for the site and for Shipow's visuals.
+
+## Manœuvre (how)
+
+Cross-comparison first, blind. Aligned the 1 Hz gig log to the audio, measured
+every track for loudness / spectrum / stereo width, then read PLN's notes
+against it.
+
+Then the master — from the **float take**, not the shipped stems, because the
+shipped stems turned out to be destroyed. Clean stems at a uniform −15 dB, sum,
+glue, loudnorm, PLN's edits (head to 0:06, encore gap trimmed, end on the
+REVOLUTION vocal with a reverb tail), and a phase-correct low-band dip on the
+one saturation event.
+
+## Prise (findings / artifacts)
+
+- **11 of 12 shipped s16 stems were hard clipped** — d4 with 2,066,873 samples
+  pinned at full scale. The 32-bit float take underneath was fine (over-0 is
+  headroom there, worst orbit +9.01 dB). A uniform −15 dB put the hottest at
+  −5.77 dBFS with **flat factor 0.000000** on all twelve.
+- Masters: streaming **−14.1 LUFS / −0.9 dBTP / LRA 9.4 / 48 kHz 24-bit**, club
+  −10.0, a 24-bit archive, and 15 tagged per-track FLACs.
+- PLN's "not loud enough — I had to set max volume" measured **−18.2 LUFS**,
+  4.2 dB under target. Ear and meter agreed exactly.
+- His "MAFIA bass goes too high for a few secs" was real and **at 1:02:10, not
+  the 1:03:10 he filed** — +24.5 dB over local median in 60–250 Hz. Found only
+  with a band-limited lens; broadband RMS saw nothing. That became a new tool:
+  `tidal-ears master saturation --band`.
+- **The rig held**: 17 xruns in the whole 1h27, in two events, despite the
+  package hitting 100 °C. And the set was played **15 for 15 in the exact
+  planned order**.
+- Two silent bugs in our own mastering tool, found the same afternoon: every
+  master it had ever rendered was **192 kHz** (loudnorm leaves the graph
+  oversampled), and every split track's FLAC header **claimed to be the whole
+  1h19 set** (`-c copy` carries the source STREAMINFO).
+
+## Sel (the shareable learning)
+
+Three times today something *sounded* fine and the metadata lied — clipped stems
+that played, a 192 kHz master nobody chose, fifteen track files each claiming to
+be 79 minutes long. None of it errored. **The failure mode of an audio pipeline
+is not a crash, it is a file that opens.**
+
+And the counterweight: the machine spent an hour hunting a "lingering arp",
+narrowed it to five orbits, and built an elegant theory around d12 — declared by
+the outgoing track, absent from the incoming one, re-declared by the next, which
+explained the symptom perfectly. d12 measured **−72.6 dBFS**. Inaudible. The
+theory was fitting a signal nobody could hear. PLN listened to five clips and
+said "wait, rehearing master i dont hear any issue?"
+
+The premise was wrong, not the orbit.
+
+## Hameçon (hook)
+
+*"Your master is clipped. Not the master you're about to make — the one you
+already shipped."*