PLN: "wheres the 'run parvagues' big HUD menu button? I see perf modes, open
bridge, then the gear 4/6 up, but below open bridge, i want a 'Launch Gig' button
that opens and focuses auto the parvagues".

Two problems behind that one sentence.

DEPTH. Pulsar was reachable only through Gear ▸ Pulsar — the hot path buried
under the tool drawer. It is now a top-level entry directly below Open Bridge,
labelled with what the click will actually do ("focus ParVagues" when it is up,
"open ParVagues" when it is not), because "Focus" vs "Open" is the difference
between a confident click and one that looks like a no-op.

FOCUS DID NOT EXIST. launchers.py's docstring said "on Wayland there's no
portable window-focus", so clicking a running app answered with the string
"already-running" and did nothing. That is true of the X11 toolbox — xdotool,
wmctrl and kdotool are all absent here and the first two cannot work under
Wayland anyway — but it is not true of THIS desktop. Plasma exposes KWin's
scripting engine on the session bus, and a script running inside the compositor
is simply allowed to activate a window.

So focus_window() registers a tiny KWin script, runs it, and unregisters it. It
matches on resourceClass (the app id), not the window title: the title changes
with every file PLN opens, and matching it would break the moment a filename
stopped containing "Tidal". The script name carries our pid so two clicks cannot
collide on an already-registered name. Verified live against the running Pulsar:
focus_window('pulsar') -> True.

Consequence worth noting: running native apps in the Gear menu used to be greyed
out, because the only honest thing a click could do was refuse. They are live
again — clicking a running app now raises it, which is the thing you actually
want mid-set.

Deliberately editor-only. Launch Gig does not start SuperCollider, boot Tidal or
arm Ardour: #116 is the standing rule that the hot path must not be able to start
the sound by accident, and one wrong click before a set is a stuck scsynth rather
than a convenience.

Every failure path falls back to reporting. A focus button that quietly does
nothing is a papercut; one that throws during a gig is not acceptable.

Bridge suite: 43 passed.

The KRKs go home after tomorrow's dawn rehearsal, so this is the last run with
real monitors — and desire / REVOLUTION / electric_hammer have not been played
this cycle at all. Read them cold so his ears go to the music.

Two findings worth the read, both the same shape as Sunday's do_it_right bug:

desire declares d4 THREE times in one block (lines 50, 56, 62), under his own
"-- FIXME BASS!!!". Later wins, so the two takes he tuned are dead code. The
winner differs in ways that do not read as a choice: it has no lpf on ^32, so
that knob is dead on the bass, and it carries no gain at all where the other two
say 1.15 — which under SuperDirt's quartic law (amp = gain^4) is 4.8 dB quieter
than the level he dialled.

electric_hammer declares d5 twice (32, 46). The superfreak voice wins, so ^89
and ^33 have nothing behind them, ^57 does something other than what the dead
copy above it describes, and hammer:7 — the sample the closing track is named
after — never sounds.

Also: the five closing seams with their ghost orbits, paste-ready orphan guards
for the two worst, his two vague_de_crime TODOs with the gain converted to dB,
and the gate result (GO, 12/12).

One methodology note recorded in the doc: silent-eval must be run --seeded here.
Unseeded it reports 12 SILENT orbits across these three tracks, every one of them
a midiOn gate sitting at default-off. That is the harness, not the music.

Four tracks carried uncommitted edits after the KRK run. They are his, they
are tuned by ear, and they were sitting only in the working tree with a
rehearsal scheduled for dawn — exactly the state that let a stale Pulsar
buffer eat committed work twice before (#95). Committing them so tomorrow's
run starts from disk, not from whatever a buffer happens to remember.

What he changed, and why it reads as deliberate:

- gimme_acid  d5 gains a `^56` jungle_breaks:15 swap and the slice pattern
  moves under `whenmod 16 8`, so the chop only bites for half of every 16
  cycles instead of running flat the whole time. d11 gets `# gain 1.5` and a
  renamed comment ("BOOM, BOOM, BOOM").
- perfect     comment-only: d1 "Famous Clubkick", d2 "Clap Mechamment
  Celebre". Naming the kick and the clap is how he finds them mid-set.
- bombe_dj    `resetCycles` commented out. This is the SET OPENER after
  quand_on_decolle; resetting the clock at the top of track 2 yanks every
  still-ringing orbit's phase from the track before it.
- vague_de_crime  d1 kick +0.3 gain, d6 gains `midiOn "^90" (ply 2)`, and two
  TODOs written in his own words: the kick wants comparing against the rest
  of the set, and d3 is "less solid than in do it right or drops or ouais".

The two TODOs are ear-work for the last-third rehearsal, not defects to fix
blind — logged rather than acted on.

Master: 61.0 min, 2 ch, 48 kHz, 24-bit. I = -14.0 LUFS on the nose, peak
-0.9 dBFS under a -1.0 ceiling, LRA 12.7 LU. L-R difference RMS is -22.4 dB, so
it is genuinely stereo and not dual-mono — worth measuring rather than assuming,
given the first attempt collapsed to one channel.

THE BUG THAT WOULD HAVE SHIPPED. take-master trims leading silence before
loudnorm sees it, so the master starts 10.94 s later in the music than the raw
take does. The segments were computed against the RAW take. Splitting with
unshifted offsets puts every boundary eleven seconds late — each track beginning
eleven seconds into the previous one's tail. The output would have played
perfectly and been wrong throughout.

Two guards in take-segments:
  --shift   subtract the measured head trim. Clamped at 0, because the first
            track legitimately goes negative: PLN focuses a .tidal a few seconds
            BEFORE he evaluates it (log says 7.4 s, first sound is at 10.94 s).
  --master  probe the mastered file and clamp to its duration. The master is
            trimmed at the TAIL too (1.45 s), so without this the last segment
            ran past EOF and ffmpeg would have written a short final track in
            silence.
Now: contiguous, starts at 0, sums to 3661.2 s — exactly the master's length.

AND A SHADOWING BUG FOUND WHILE FIXING IT. The per-segment loop bound `start`
and `end`, the same names as the take window, so the summary line printed the
LAST SEGMENT as if it were the whole take: "1970-01-01 01:55 — 5.1 min". The
segments were correct the whole time; only the report lied. That is the kind of
defect that ships, because the payload looks right and nobody re-reads a header.

GENERATIVE METADATA, per PLN's ask, derived not typed:
  title    backlog.md via setlist.py (the SSOT reaches the ID3 tag)
  TBPM     backlog BPM, NOT setcps — reference_tempo_from_audio, setcps is not
           felt tempo. gimme_acid is setcps 80 played as 160 half-time DnB, and
           you_my_sunshine is setcps 144 against a felt 166. Tagging 80 would
           have been a measured number that misleads.
  grouping the set section (Ouverture / SUNSET / NuJazz / NUIT)
  comment  orbit inventory from the .tidal via pvlint's parser, plus the source
           path, so any track can be traced back to the code that made it

Metadata says REHEARSAL, not gig: album "OPAL 2026 — rehearsal, J-6", date
2026-08-02, no venue, no lineup, no poster (feedback_metadata_vs_mastering — gig
metadata is canonical only in Web/www/next/content/lives/).

lime_tacos is track 7 at 1:49. He played it, then cut it from the set an hour
later. The take records what happened, so it stays, tagged "not in the final
set".

Committed: segments.json, track_metadata.json, analysis.json — the provenance.
Gitignored: the FLACs and MP3s, ~1.3 GB of derived audio regenerable from the
Ardour stems in two commands.

`tidal-ears master mix` has the right recipe and this reuses it verbatim. What
it does not know is how Ardour lays out a take, and that produced two defects
that only appear on this material — both of which I only caught by inspecting
the output file rather than trusting "Done:".

  1. MONO. Ardour writes each orbit as TWO mono files, `Take95_Tidal 04-1%L.wav`
     and `%R.wav`. Globbing all 24 into amix sums them into ONE channel. The
     stereo image is gone — and ParVagues pans deliberately (`# pan 0.8` on the
     hats, `0.2` on take_5's sleepwalker), so that is not a technicality, it
     discards a compositional layer silently. Each pair is now `join`ed into
     stereo first and the mix is 12 stereo streams.

  2. 192 kHz. loudnorm resamples internally to 192 kHz to measure true peak, and
     with no explicit output rate ffmpeg keeps it. A 48 kHz session produced a
     624 MB 192 kHz file carrying no extra information. `-ar` is now pinned to
     the source rate. NB this one affects every master the sister tool has made.

UNIFORM TRIM, not per-stem normalisation. gain_for_stem normalises each stem
toward a target peak, which is right for stems of unknown provenance and wrong
for these: PLN balanced these twelve orbits by hand, on faders, while playing.
Normalising individually rewrites that balance — measured at 4.0 dB of shift on
Take95, mostly making the hats hotter. One trim for every stem preserves the mix
he played; loudnorm sets the absolute level. Same conclusion take-lens reached.

AND THE MEASUREMENT ITSELF HAD A TRAP. The first version used `volumedetect`,
which reports max_volume against a fixed-point full scale and SATURATES at 0.0
dB. Every one of these float stems read as exactly +0.0, so the derived trim was
-6 dB when the truth needed -19.3. Twelve identical +0.0 readings is not data,
it is a clipped instrument — `astats Peak_level` reports the real values and
they match the sister tool's analyze pass exactly (d4 +13.3, d3 -3.3).

Refuses to render if an orbit has only one side, rather than panning it hard.
Verifies channels and rate of its own output and exits non-zero if either is
wrong — the check that would have caught both original defects.

output/Take95/segments.json — 11 tracks over 61.2 min, derived from gig-log by
tools/take-segments.py rather than detected in the audio. Committed because it
is the provenance for the album's track boundaries and titles: without it, the
split is an unexplainable list of timestamps.

Titles come from backlog.md via setlist.py, so they are the same strings PLN
wrote, not a second copy that can drift.

  1. bombe_dj   3:53   |  2. wap        3:19  |  3. do_it_right  8:25
  4. take_5    10:19   |  5. piment     4:30  |  6. ouais_je_funk 4:54
  7. lime_tacos 1:49   |  8. perfect    6:52  |  9. gimme_acid    5:19
 10. vague      6:47   | 11. sunshine   5:04

lime_tacos is in the take at 1:49 — he played it, then cut it from the set an
hour later ('actually weak'). The take is a record of what happened, not of what
the set became, so it stays.

Third session running, he raised faders by hand and hit Ctrl+S, then asked the
right question:

    "ok raised and saved 10 and 12, saved, check again.
     but ideally we should not depend in that?"

He is right. The gate CAUGHT the problem all three times — the dependency was
never on the check, it was on him having hands free to fix it. On a Saturday at
19:00 in a field that is not a dependency you want.

THE MECHANISM, now understood: the Ardour session IS the boot state. He ends a
set with the faders pulled down, which is what you do at the end of a set,
Ardour saves that, and the next launch comes up silent. Today's save had NINE of
twelve at -inf, and all 15 setlist tracks would have lost an orbit — d10 carries
the riser convention (#33), so all ten risers would have been silent and he
would never have heard them go.

tools/fader-baseline.py: --capture / --check / --restore.

WHY THIS IS ALLOWED TO WRITE FADERS when the standing rule says never to. The
rule exists for three reasons — it desyncs the file from the physical desk,
CC77 down is total silence, and levels are MUSICAL INTENT an agent must not
invent. All three are about INVENTING a level on a LIVE desk. This does neither:
it replays a value PLN set and approved (stamped with date and source session),
it refuses to write under a live session so the desk always wins, and it can
only address routes named "Tidal NN" — Master, monitor and busses are not
reachable by this tool at all. It writes a .bak first, because a corrupted
performance session six days out is not a recoverable mistake.

--capture REFUSES if any fader is silent. Capturing end-of-set silence as the
baseline would enshrine the exact bug this undoes.

Baseline captured while the mix was green: 12 faders, 01 at 0.0 dB through 09/12
at +6.0 dB.

PROVEN BY MUTATION on a COPY, never his live session: pull the same nine faders
to 0.0, --check reports all nine with their baseline deltas and exits 1,
--restore puts them back, and check-mix.py independently agrees every orbit
reaches master again.

AND THE TEST FOUND A REAL BUG IN THE GUARD. The first version asked "is ANY
Ardour running" — the safe-LOOKING answer and the wrong one, because it refuses
to repair a session nobody has open, and it made the restore path untestable
while his live session was up. The hazard is writing UNDER a live session, so
the guard now matches on the resolved session path. Still refuses his open one;
no longer refuses everything.

Wired in two places: a HARD "fader baseline" check in the gate, and a --converge
action that restores automatically when Ardour is closed. gig-up: GO, and
--converge is a clean no-op on a healthy rig.

Last KRK day. The chart PLN asked for, and the honest reading of it: the SET is
finished and derived rather than maintained, the TOOLING is done and should stop
there, and the long pole is REHEARSAL — 11 of 14 tracks played, three never
heard this cycle, and those three are the FINALE and ENCORE.

gig-up is NO-GO on exactly one thing and it is not code: nine of twelve Ardour
faders sit at -inf in the saved session (#124, third recurrence). Mechanism now
suspected — he ends a set with faders down and that is what Ardour saves, so the
saved session is a snapshot of end-of-set silence. Thirty seconds to fix today,
but #20's template is the real answer before Saturday.

Pacing measured for the first time rather than guessed: Take95 ran 61.2 min for
11 tracks = 5:34 each, so 14 tracks is ~78 min against a 90 min slot. Twelve
minutes of slack.

Also carries the buffer-clobber warning for tomorrow's editing session: close
and reopen any tab open yesterday, because Window: Reload does not do it.

Today's buffer clobber should have been caught. It was not, for two reasons that
are each worse than the bug:

1. tools/check-drift.sh was NOT wired into gig-up.sh at all. Grepping this file
   for it returns nothing. The memory note claiming it was wired in was simply
   wrong, and had been trusted since.
2. The script was mode 100644, so invoking it exits 126 (permission denied) —
   the FOURTH occurrence of the chmod-vs-git-mode bug, and the reason the "tools
   executable" check exists. check-drift.sh was not in that check's list either.

A check that exists, is documented as guarding the gate, and is never called is
worse than no check: it buys confidence it does not deliver. That is the same
lesson as the two dead checks found when gig-up was first assembled.

NEW HARD SECTION 1c "surface grid intact", asserting two things:
  * check-drift.sh passes (the committed grid still holds)
  * migrate-columns.py --plan reports 0 moves (the WORKING TREE needs none)

The second is what actually caught today's clobber. When Pulsar wrote its stale
buffer over wap.tidal, d4's octer knobs reverted ^32 -> ^52, undoing a migration
committed and pushed hours earlier. That was invisible in `git status` — the
file just reads as modified, mixed in with PLN's own real edits. What made it
visible was migrate-columns being IDEMPOTENT: 0 moves one minute, 1 move the
next. Any property a tool can re-derive is a property a gate can assert.

PROVEN BY MUTATION, and the first mutation was wrong in a useful way. Flipping
ONE knob ^32 -> ^52 did not trip the check, and that is correct: B4 and C4 are
both legal d4 slots, so two effects merely swapping slots is not drift. The real
clobber put BOTH effects on ^52 — one knob driving two effects — and that trips
it immediately, printing the offending moves. Had I stopped at the green run I
would have shipped a check I believed in for the wrong reason.

Also adds check-drift.sh and take-segments.py to the executable-bit check, and
records 100755 for both in git.

Splitting a livecoded set by listening to it is hard, and we have the scars for
it: orbits stay up across a transition, there is no silence at the seam, and a
gap-based detector finds the breakdowns instead of the track changes.

But gig-log has been writing the answer at 1 Hz this whole time. Every time the
active .tidal file changes in Pulsar it records

    {"k":"track", "path":"live/collab/raph/mafia_sans_serif.tidal",
     "from":"live/midi/nova/dnb/liquid/you_my_sunshine.tidal"}

which is ground truth for WHAT was playing, for free, with no DSP. The audio
never had to be interrogated at all.

THE ONE COMPLICATION is that PLN tabs around while he plays — glancing at the
next track, checking a helper, flicking back. Take95 contains
wap -> do_it_right -> wap -> do_it_right inside 90 seconds. So a segment
shorter than --min-dwell is a GLANCE and gets absorbed into whichever NEIGHBOUR
is longer (the track you were on, not the one you peeked at).

Debouncing needs a fixed-point loop, not one pass: dropping a glance can leave
two spans of the same track adjacent, merging those can leave a span that is
still under threshold. Take95's take_5_drops/do_it_right flicker takes three
rounds to settle. 29 raw events -> 11 tracks.

Window comes from the AUDIO (soxi duration + mtime of the stems), not from the
log, so a recorder still running after the transport stopped cannot stretch the
take. Titles come from tools/setlist.py, so a rename in backlog.md reaches the
ID3 tag — the SSOT rule now extends all the way to the export (#123).

Emits the --segments shape `tidal-ears master split` already accepts, so the
mastering pipeline needs no changes.

Take95 (2026-08-02, 61.2 min): 11 tracks, bombe_dj through you_my_sunshine —
which is exactly where PLN said he stopped ("didnt play desire/revolution/
hammer").

PLN: "in you my sunshine the bass easily saturates too, not sure if code or
ardour setup". Code, measurably — and the same number as this morning's kick.

d4 is not one bass, it is two playing at once: a `superimpose (const ...)`
bassWarsaw plus the chopped no_sunshine loop. ^32 crossfades between them:

    superimpose   # gain (range 0.5 2   "^32")   ->  peak +24.1 dB at ^32 max
    main          # gain (1.8 - 0.6 *   "^32")   ->  +20.4 dB at ^32 UNTOUCHED

Both are crushed off the same ^52. So the orbit is loudest at BOTH ends of the
crossfade and its default resting state is +20 dB — a fader ride cannot rescue
that, because the shape is wrong, not the level.

Now `range 0.5 1.5` and `1.5 - 0.3 * ^32`: both ends land at +14.1 dB, so the
crossfade holds roughly constant loudness instead of bulging. This is
[[reference_superdirt_gain_is_quartic]] applied a second time in one day — the
gain^4 exponent is why 2.0 reads harmless and meters at +24.

Also: "d3 should also be the classic techno runner companion on 43, so it
hits(4,8,1)". ^43 was a bare `ply 2`. Now `ply 2 . struct "t(4,8,1)"` — struct
puts the hats ON the euclid rather than thinning what is already there (mask can
only remove), and ply doubles each into a running 8. Released, the liquid
pattern returns.

REAPPLIED after a buffer clobber: wap's d4 octer/octersubsub had reverted from
^32 back to ^52 when Pulsar wrote its stale buffer on close. Exactly
[[reference_pulsar_buffer_clobber]], and exactly what #95 exists to prevent.
Caught because migrate-columns --plan is idempotent and went from 0 moves back
to 1. PLN's own new FIXME on that line ("too wobbly... not fan of the 44/76
slicing") is preserved.

SET REORDERED by PLN: gimme_acid -> mafia -> sunshine -> vague -> desire.
"sunshine great cinematic ending, but sunshine-to-vague-to-desire makes more
sense imo." vague->desire is now a CLEAN transition, and sunshine's 166->120
drop into vague is the one seam to watch. Setlist and preload regenerated from
backlog.md; 52 banks cover all 15 tracks.

PLN added quand_on_decolle under a new ## SOUNDCHECK heading mid-session. The
gate caught both downstream consequences without being asked:

  setlist vs backlog  FAIL -> the .txt was 14, backlog.md was 15
  preload covers set  FAIL -> 'airports' would have been read from DISK on
                              first play, i.e. crackle, on the soundcheck track

That is the #120 failure mode caught BEFORE the venue instead of at it, and it
is the second time in two days the backlog->setlist->preload pipe has earned
itself. gig-up: GO, 10/10 cold.

preload.scd takes effect at the next SuperDirt boot — NOT applied live, since
scsynth is up and restarting it costs ~15s of silence.

Densest ear-feedback session logged so far, and the finding that makes it worth
re-reading: not one of PLN's 'it sounds wrong' reactions was a mix problem.

  'muffed'      -> a SynthDef's own 1 kHz Moog lowpass, no cut group, shared bus
  'loud'        -> SuperDirt amplitude is gain^4, so `# gain 2` is +24.1 dB
  'irritating'  -> a literal ascending scale being used as a melody
  'no buttons'  -> CCs that were never mapped
  'weak'        -> the one track the column remap never reached

Includes the gain->dB table and the measured claim that the OPAL set's median
orbit sits at +11.7 dB while four sit at +24.1 dB — and the rule that follows:
when an orbit clips, rank the track's orbits in dB before touching a fader. A
10 dB intra-track outlier is a bug; 2 dB is a mix choice. Feeds #91 and #17.

PLN, mid-run: "i hear no d7? or its muffed like '# att 0.2' or sth?" — then,
after experimenting: "tried adding 'arp up' before the chord, its now a set of 4
muffed short notes LOL. but d7 still muffed".

Read the SynthDef instead of guessing the synth (CLAUDE.md: never infer a
sound's role from its name, validate). SCLOrkSynths/SynthDefs/bass/moogBass.scd:

    arg ... cutoff = 1000, gain = 2.0, att, dec, sus, rel, chorus = 0.7;
    filter = MoogFF.ar(in: osc, freq: cutoff * (1.0 + (0.5*filterenv)), gain: gain)

A 4-pole Moog lowpass parked at 1 kHz, on a SynthDef whose own metadata reads
`category: \bass, tags: [\pitched, \bass]`. Every harmonic above 1 kHz is gone at
24 dB/oct. A "CHOEUR GENERATIF" voiced on a bass synth cannot be anything but
mud, and no amount of gain, room or crush was ever going to fix it. `# cutoff
3500` opens it — 2500 warm pad, 3500 choir, 6000 buzzy.

TWO TRAPS in this SynthDef, now written into the file where the next reader
will hit them:
  * it DECLARES att/dec/sus/rel and then ignores them — the envelope is a
    hardcoded Env.adsr(0.001, 0.3, 0.9, 0.2, amp). That is why PLN's `# sus 2`
    did nothing and he commented it out. Shape this voice with legato.
  * `gain` in this SynthDef is the MoogFF filter RESONANCE, not a level.

Three more real defects in the same orbit, all found by tooling, all fixed:
  * legato 2 -> 1.2. His new arp makes quarter-length notes; 2x overlap smeared
    them into each other. That is the "4 muffed short notes".
  * d7 had NO cut group at all, so held voices piled up — on a synth that also
    carries its own 7-tap chorus. Now `cut 7`.
  * crushbus restored, on slot 71 rather than d4's 41. The shared-bus case that
    became PV011 twenty minutes ago.

And PV004 caught one PLN had not mentioned: d11 (THE FLUTIST) and d12 (THE
SLEEPWALKER) both sat on `cut 11`, truncating each other instead of layering.
d12 now owns cut 12.

Also in this commit: mafia_sans_serif's d7, whose diagnosis is written up in
b8344ac1. take_5_drops now 0 errors / 0 warnings, silent-eval --seeded ok.

THE PHANTOM. PLN drafts in backlog.md, and he drafts in LIST ITEMS. Mid-run he
wrote himself a deliberation:

  -- Q: This intro is quite raptech. jazz intro? quand on decolle opening? reflect

and quand_on_decolle silently joined the OPAL set. The resolver's "codename
carries extra words" branch is a subset test, so {quand, on, decolle} being 3 of
those 13 words was a match. It then flowed downstream exactly as designed — into
the .txt, the preload plan, the remap scope and the gig gate. The pipe worked
perfectly, on garbage.

Two guards, because either alone is too clever:
  * _NOTE_RE skips items marked Q:/NB:/NOTE:/TODO:/FIXME:/IDEA:, and PRINTS what
    it skipped. A note dropped silently and a track dropped silently look
    identical from inside the parser, and only one of them is survivable.
  * the extra-words branch is bounded to 4 decoration words. Real entries spend
    at most 4 ("[129] Desire [TODO BASS ETC!]"); that note spent 10.

An UNMARKED prose line now RAISES instead of resolving. That is deliberate and
tested: red gate, PLN adds "Q:" or deletes the line. A phantom track goes to the
venue. Loud beats clever.

(He has since re-added quand_on_decolle under an explicit ## SOUNDCHECK heading,
as a real entry. It arrived carrying the full off-by-one-column corpus pattern —
11 moves — now migrated. That is the pipe working on something real.)

DEAD BUTTONS. PLN, playing: "not seeing effects on button 44 nor 43?" on
do_it_right, "no buttons 42 43?" on ouais_je_funk. They were not broken, they
were never mapped. Now:
  do_it_right   ^43  d3 hats — HIS OWN TODO implemented verbatim: "an opening
                     legato or att to make it slowly go from hard dry h to an
                     open ho? as in take 5 drops". Held, legato walks 0.4->2.5
                     and hpf 7k->2.5k over 8 cycles; released, it snaps back.
  do_it_right   ^44  d4 bass — octave down, DROP TO THE CELLAR
  ouais_je_funk ^42  d2 clap — the clap had no gate at all. Handclap roll.

d4's crush/octersub land on ^52/^32 exactly as PLN asked. That required
REORDERING the two lines, not renumbering them: migrate-columns assigns knobs in
order of declaration, so the file now carries a comment saying the order is
load-bearing. Fighting the tool by hand would just be undone next run.

Set: 15 tracks, 0 moves pending, silent-eval --seeded 15/15, pvlint 0 errors,
setlist tests all green (now pinning lime_tacos OUT and the phantom absent).

PLN, at rehearsal, on take_5_drops' d7 coming out muffled:

  "i saw also its crushbus was 41, a d7 with sub 4 is sus -- and two effects on
   same bus is a redflag as it means fighting for bus! so we gotta flag warning
   any 'bus nXY on dX' as sus (e.g. bus 41 (d4's) on d7 here), and any shared
   bus (^41 on d4 and d7) as PV error"

Two rules, exactly as specified.

PV014 (warning) — a bus slot numbered for a different orbit. The convention is
<orbit><slot>: d4 owns 41/42, d10 owns 101/102, and a bare `bus 4` on d4 is the
same idea with one slot. MEASURED before encoding it rather than assumed: 894
of 984 bus references under live/ conform, 90.9%. A real convention, not a
tidiness preference. Mis-numbering is *how you accidentally share* a bus, so
this rule sits upstream of PV011 and catches the cause, not the symptom.

PV011 promoted warning -> error. It shipped as a warning on the theory that
sharing a bus is legitimate glue. Rehearsal disproved that theory.

The asymmetry with PV008 is deliberate and now documented in both rules and
pinned by a test, so no future refactor quietly aligns them: yesterday PLN
DEMOTED a shared button to a warning ("its ok when two percs are linked"),
today he PROMOTED a shared bus to an error. A shared button is one hand doing
two things on purpose and you hear it instantly. A shared bus is two orbits
fighting over one effect instance with the winner decided by event density —
you don't hear a mistake, you hear a slightly wrong sound and blame the synth.
Surprises you can hear are warnings; surprises you cannot are errors.

THE PAYOFF, first run, in a track still in the set. mafia_sans_serif's d7 was
already carrying PLN's own note "FIXME ACID WEIRD NOT GOOD". It had THREE
collisions with d4's moog bass, none visible in the code:

    # cut 4        every acid note truncated the bass, and vice versa
    # crushbus 41  one crush instance, ^35 and ^32 fighting over it
    # octerbus 42  one octer instance, ^55 and ^52 fighting over it

d7 now owns cut 7 and slots 71/72. That is a diagnosis for a FIXME he wrote
without knowing the cause, produced by a rule written twenty minutes earlier.

Set is 0 errors / 9 warnings, silent-eval --seeded 14/14 ok, pvlint 68/68.

PLN, mid-run: 'wait bass was also on 89 not yet 76 whyyy i thought we were
systemic :((' — and he was right to be annoyed. The answer is embarrassing and
exactly the bug we already fixed one layer up.

tools/migrate-columns.py hardcoded SETLIST = armada/setlist_opal2026.txt. That
.txt was 13 tracks while backlog.md was 15. So when the remap ran, the two
tracks PLN had just added were the only two it never saw — ouais_je_funk and
lime_tacos kept their pre-remap CCs and he found out by reaching for a control
that wasn't there. Same shape as #120/#121: a generated artifact consumed as if
it were the source. resolve() now calls setlist.tracks(), i.e. backlog.md.

Re-planned against the real set: 22 pending moves across 4 tracks, 11 of 15
already clean. Then PLN cut lime_tacos ('actually weak, will have more fun on
ouais funk and perfect') — removed from backlog.md, and the 14-track setlist
plus the plan regenerated themselves from it. That is the pipe working.

Applied, 10 moves:
  wap        d4  octer/octersubsub  ^52 -> ^32 (B4)
  piment     d7  modIndex/octersub  swapped onto B7/C7 in declaration order
  ouais      d3  snare gate         ^44 -> ^43 (E3)
  ouais      d4  crush ^53->^52 (C4), octersub -> ^32 (B4)
  ouais      d5  width/crush        swapped onto B5/C5
  ouais      d5  superimpose        ^57 -> ^89 (F5, gestures live on row F)
  ouais      d7  UPDOWN             ^55 -> ^35 (B7)

d4's 'ply 2' had been authored on a KNOB (^56) and the migrator dutifully gave
it d4's B4 slot, which pushed the octersub into an overflow FIXME. But PLN had
just asked for exactly the opposite layout — 'd4 52 a crush? 32 a octersub
maybe?' — so ply moved to ^44 (E4, a gate on the gate row, free now that d3's
snare vacated it) and both effects got their proper knobs. No FIXME left.

Verified: migrate-columns --plan reports 0 moves, silent-eval --seeded 14/14 ok.

PLN typed a TODO comment into the middle of `d12 $ gF3 $ gM3`, landing the `3`
after the `--`. The line became `d12 $ gF3 $ gM`, and `gM` does not exist:

    Variable not in scope: gM :: Pattern ValueMap -> Pattern ValueMap
    Perhaps use one of these: 'gM1', 'gM2', 'gM3'

One parse error silences the WHOLE do-block, so this was not 'd12 is quiet' —
it was 'do_it_right makes no sound at all', found by silent-eval --seeded while
he was mid-run. Comment moved above the line, where it cannot swallow code.

The TODO itself is preserved verbatim: he wants more of ^18/^19 on this lead.

J-6, the one-before-last KRK rehearsal. Per-track watch list for the 1h30 run,
because ten of the fifteen tracks have something in them PLN has never heard:
seven new d10 risers, the take_5_drops d5 structure fix, ouais_je_funk's d10
finally mutable, lime_tacos' cymbals relocated to d12, and two tracks that were
not in any tool's idea of the set until yesterday.

Also carries the fresh ghost matrix (9 of 14 transitions, 18 orphans) recomputed
on the real 15-track order — and the pleasant surprise that adding the two
missing tracks produced the longest clean stretch in the set,
piment -> ouais -> lime -> perfect.

PLN: "for lime if d11 taken lets push d6 to free d12 :)" — d11 is held by
"Menace de retour", and d10 is now the safe riser, so d12 was the free slot.

The move is genuinely one token: d6 appears exactly once in the file, carries no
`cut 6` and no orbit-bound crushbus, so nothing had to follow it. ^44 stays
shared with d3 DELIBERATELY — that is the idiom PLN ruled correct: one press
re-rhythms hats and cymbals together.

Side effect worth knowing at the desk: PV008 only inspects orbits 1-8, so with
the cymbals on d12 the shared-button warning stops firing entirely. The link is
unchanged; only the lint's field of view moved. Do not read the green as the
coupling having gone away.

WATCH ON THE RUN: Ardour Tidal 12 sits at +6.0 dB and Tidal 06 at -3.8 dB, so
the cymbals land ~9.8 dB hotter than before. Ride Tidal 12 down rather than
re-gaining the pattern until PLN has heard it.

Verified: silent-eval --seeded ok (every declared orbit emits), pvlint clean.

PLN, on lime_tacos' ^44 driving both d3 (hats) and d6 (cymbales sol):

    "lime tacos its ok when two percs are linked, e.g. here, change rythm of
     both 3 and 6. [...] im fine with this, not a NOGO, just a WARN?"

He is right, and the interesting part is WHY the rule was wrong rather than that
it was. PV008 was earned on a real bug — the 2026-07-29 phase-2 remap left d1
holding ^42 after d2 was given ^42, and perfect.tidal shipped one button firing
two orbits. That bug is real and this rule is how it was found.

But the accidental case and the deliberate one are TEXTUALLY IDENTICAL. Two
orbits, one button CC, no way to tell from the source whether PLN meant it. The
rule had quietly encoded "one button drives one orbit" as an invariant of his
style, when it is only a hypothesis about it — and the hypothesis was wrong. He
groups percussion on purpose: one gesture re-rhythms the whole group.

A check that cannot distinguish intent must not hold the gig gate shut. It
reports; PLN decides. So: severity error -> warning, and the detail now states
both readings and what distinguishes them (deliberate percussion grouping vs a
control inherited from a remap) instead of assuming the second.

Effect: tools/gig-up.sh goes NO-GO -> GO. This was the last blocking failure, and
it was blocking on a correctly-written track.

Added test_pv008_is_a_warning_not_an_error, which quotes PLN's ruling, so nobody
"tightens" this back to error without meeting the argument first. pvlint 63/63.

Third time. tools/gig-up.sh's 'tools executable' check (added yesterday for exactly
this) is why it surfaced: setlist.py went into 04fc8c7c with content but mode 100644,
so a fresh clone would get 'permission denied' from a HARD gate check. chmod +x
touches the working tree; git stores the bit separately and needs
`git update-index --chmod=+x`, and even then a pathspec-limited commit can leave
HEAD at 100644 while the index says 100755 — which is what happened here.

Verify with `git ls-tree HEAD <file>`, never `ls -l`.

PLN: "can you add a safe riser on all set tracks that dont have one, and consider
moving the d10 to d9 (eg a synmenace iirc) to have d10 riser always?"

THE CONVENTION. d10 = riser, so the gesture lives on the same column in every track
and his hands stop having to remember per-track. It was already true in three tracks
without anyone naming it — do_it_right (mask "<f!4 t f!3>"), take_5_drops (the
"A FAIR RISER" comment on d10) and vague_de_crime, whose line #33 designates as THE
safe riser because a 1-in-8 mask cannot run away:

    d10 $ gF1 $ gM3 $ mask "<t f!7>" $ "risers:8"

Copied verbatim into the 7 tracks where d10 was free: bombe_dj, wap, lime_tacos,
you_my_sunshine, mafia_sans_serif, desire, electric_hammer. Inserted contiguous with
the existing d-blocks, because all three pre-existing risers live inside the main
block rather than standalone — a riser in its own block would need a separate
ctrl+enter, which is not the gesture. 10 of 15 tracks now have it.

WHY THE RELOCATION IS NOT IN THIS COMMIT. "Consider" did real work here. The two
tracks PLN was remembering are piment_bresilien (d10 = Menace, d9 free) and
ouais_je_funk (d10 = SynMenace, only d12 free) — but moving an orbit is not a
one-line move, because orbit N is simultaneously the LCXL column, the Ardour track
and the cut/crushbus group:

    piment's d10 block carries  -- A6 = d10's effect  on its crushbus CC ^18
    d9's fx CC is ^17, d12's is ^20      -> the CC must move with the sound
    Ardour Tidal 10 = -2.2 dB, Tidal 09 = +6.0 dB, Tidal 12 = +6.0 dB
                                         -> the synth lands +8.2 dB louder

And the argument that settles it: perfect, gimme_acid and the_revolution declare all
12 orbits, so d10 is never free there and the convention caps at 12/15 no matter what.
Relocating buys 2 tracks for two CC remaps, a cut-group change and two synths jumping
+8.2 dB — on material PLN has not rehearsed since, 7 days out. Deferred by choice, not
forgotten; his call to overrule.

#99, d5 of take_5_drops: `>|` takes structure from the RIGHT, so `slow 4` note
patterns were REPLACING the slice rhythm with one event per four cycles — the
"steals structure" symptom, confirmed. Now `|>`, which keeps the left structure and
takes only the note value. Note line 35 immediately above was already doing it that
way, so this makes the block self-consistent rather than inventing an idiom. Marked
TODO @pln verify as asked.

Also ouais_je_funk d10 read `gF3 $ gF3` — the same helper twice, so the orbit had no
family mute and NOTHING on the surface could drop it (pvlint PV013). Now `gF3 $ gM3`.
Also TODO @pln verify.

VALIDATION: all 15 setlist tracks still compile (gig-up "setlist compiles" green,
silent-eval cold-queries every dN), preload still covers the set at 53/53 banks, and
pvlint added zero new findings — still the one pre-existing error, lime_tacos' ^44
driving both d3 and d6. That one is left alone deliberately: ^44 is neither orbit's
home gate (d3 = ^43, d6 = ^58/^90), so fixing it means rebinding a button PLN's hands
already know, which is a taste call and not lint's to make.

PLN's ruling yesterday: "set SSOT is the backlog entry, always. if in the backlog
i add a track or reorder them, its the truth." Then today, the part that decided
the design: "if i added new ones, id not need to tell you to update, it should
flow downstream, ditto if remove one."

So this is not a resolution, it is a pipe.

WHAT WAS ACTUALLY WRONG. Applying the ruling for the first time found the backlog
listing 15 tracks while every tool in the repo believed 13:

    missing everywhere:  ouais_je_funk (120)   lime_tacos (110)
    wrong order:         backlog has you_my_sunshine (166) BEFORE mafia_sans_serif

ouais_je_funk was not a stale wishlist entry — PLN committed a kick rework to it
in 23f60e76, yesterday. And the drift had already cost him something audible:

    preload warmed   50 banks   (13 tracks)
    the set needs    53 banks   (15 tracks)
    would cold-load  dr  gfunk_bass  h2ogmcy

Three banks read from disk on first play. That is the same mechanism PLN heard as
crackle on the take_5_drops transition (#120) — a bug we thought was fixed, still
live, just smaller. check-preload.sh reported "ok" the whole time, because it asked
set-coherence, which was one of the wrong answers. The freshness check had
inherited the staleness it was built to catch.

ROOT CAUSE: eight tools each hardcoding `SETLIST = armada/setlist_opal2026.txt`,
and one human keeping that file in sync by hand.

THE SHAPE, and why it is not "make all eight import a module". The .txt stays
exactly where it is, same format, so all eight consumers are untouched — this is
J-7 to a gig and a refactor that breaks a tool costs more than it buys. What
changes is its STATUS: generated, never hand-edited, with a HARD gig-up check
that it still matches the backlog. One parser, one artifact, one answer at a time,
and a gate that fails loudly the moment they drift.

  tools/setlist.py         the parser. Resolves PLN's freeform codenames
                           ("Ceci n'est pas Une Bombe" -> bombe_dj, "Le shifteur
                           marteau ? to finish!" -> electric_hammer) via aliases +
                           token containment, and raises rather than guessing.
  gig-up --converge        regenerates the setlist BEFORE preload. That order is
                           load-bearing: preload reads the setlist, so doing it the
                           other way warms the old set.
  gig-up "setlist vs backlog"   HARD, and placed before every check that reads the
                           setlist — a stale list makes "setlist compiles",
                           "preload covers set" and "transition ghosts" all answer
                           about a set PLN is not playing.

WHAT THE END-OF-SECTION DETECTION TAUGHT ME. First attempt whitelisted the
headings that count as part of the set ({FINALE, Encore, ...}) and died on line
one against "## Livecoding Techno DNB Nu-jazz". PLN's headings are prose; any
whitelist is a list of the ones he happened to write that day. The working rule
ignores heading semantics entirely: run to the next top-level heading, take the
list items. Non-track content does not survive that filter — "> Sunset Forest",
"1h30 - 20h30-22h", "??TRANSITION???" and the trailing url: block are none of
them list items.

VALIDATION, and specifically not "it returns the 15 tracks currently in the
backlog" — a hardcoded list passes that too. tools/tests/test_setlist.py MUTATES
a synthetic backlog and asserts the output moved: add a track (appears, in PLN's
position, not appended), remove one (disappears), swap two lines (output swaps),
unresolvable name / duplicate / missing section (all raise). 18/18.
Then the gate itself, mutated: dropped a line from the .txt -> caught; swapped
two lines, same membership -> caught as "same tracks, DIFFERENT ORDER".

AND IT PAID OFF IMMEDIATELY. With the real 15 in the setlist, pvlint went red on
two tracks that had been invisible to every tool until today: an unmutable orbit
in ouais_je_funk (gF with no gM3 — nothing on the surface can drop it) and button
^44 driving both d3 and d6 in lime_tacos (one press, two orbits). Neither is new
breakage; both were simply never looked at. Left for PLN — they are musical calls.

The 53 banks are warm on the rig now: 54/54 banks OK in 7.0 s.

Reworking d4's texture: the lpf range sweep comes out (resonance stays), crushbus
goes 16->3.5 on ^52, and room 0.4 / sz 0.5 / dry 1.1 opens the tail.

Three closed tasks from the 2026-08-01 session, written for a cold reader. The
through-line worth mining later: the rig SENSES everything and ACTUATES nothing.
scsynth died and sclang logged it, the Bridge warned 63 times, gig-up saw it and
printed GO — and separately, PLN's ears found a preload bug that four green tools
had missed. Plus the session's sharpest lesson: two of my own fixes (a watchdog, a
chmod) were themselves broken in exactly the way they were meant to catch.

PLN's own rehearsal edits, 2026-08-01, saved before a break. Committed as their own
human commit so nothing is lost — IN PROGRESS, not a finished arrangement.

* ouais_je_funk — the kick becomes gate-driven instead of static:
  "[jazz,kick:5]*4" -> midiOn "^41" / midiOff "^41" variants over a layered
  "[techno:0,808bd:2,909,kick:4]" with gain 1.5 # lpf 400. Old "kick:5" kept
  commented rather than deleted.
  Also: the lead drops an octave — note "<b b b g>@7 <cs6!3 fs6>" is now
  ("..." - 12); the superimpose gate moves ^18 -> ^20; crushbus 101 added on ^18;
  and both d11 arpeggio blocks (pinkyup/thumbupdown, diverge/disconverge) are
  commented out rather than removed.
* desire — crushbus 41 (range 16 2.5 "^52").

Left as PLN saved it. He resumes rehearsing Saturday afternoon.

--converge is the CLI big red button: bring the rig to ready, then let the existing
checks prove it. Scope is SERVICES, not apps — it starts systemd units it owns and
REPORTS Pulsar/Ardour with the command instead of launching them, because a script
that opens windows under someone is a script they stop trusting (and Ardour has a
session + crash-recovery dialog, #20). Launching apps belongs to the Bridge (#116).

Safe at any moment: every action is conditional on the thing being broken, so a
healthy rig converges to a no-op — verified, scsynth PID unchanged. It handles the
exact 2026-08-01 failure explicitly: unit ACTIVE but audio server DEAD needs a
RESTART, since `start` is a no-op on an active unit. Distinguishing those two is the
whole lesson. Preload is fixed BEFORE SuperDirt boots, or the warm would not happen
until the next restart. And it waits for scsynth to appear rather than guessing —
guessing is what made the first watchdog flap.

=== THE THIRD TIME, AND THIS TIME IT WAS US ===

gig-up's first run (1219c704) found check-boot.sh and check-tracks.sh at mode 644 —
never runnable as `tools/check-tracks.sh`. That fix turned out to be FAKE: the chmod
only ever touched the working tree. Git has recorded 100644 for both from that day to
this, so every fresh clone still got a broken pre-gig gate.

And gig-up.sh had the same bug, from birth. Committed 100644 by 1219c704 and again by
6ab09eaa. It ran only because the local working tree happened to carry the bit; a
`git checkout master` materialised it at 644 and it died with "permission denied".

That failure was nearly invisible, which is the real story: the caller filtered its
output through sed, so a run that never executed LOOKED like a clean no-op, and I
reported "no-op confirmed" about a script that had not run. A false green produced by
the absence of a program.

So the gate now checks its own hands — "tools executable", HARD, verifying BOTH the
on-disk bit and what git records. It found four real problems on its first run:
    disk=x  git=100644  check-boot.sh          <- day-one "fix" never committed
    disk=x  git=100644  check-tracks.sh        <- same
    disk=-  git=100644  sc-watchdog.sh         <- not executable AT ALL
    disk=-  git=100644  tests/test-sc-watchdog.sh
The third is the one that matters: parvagues-sc-watchdog.service had been ENABLED AT
BOOT with an ExecStart it could not execute. It was stuck `activating`, restart-
looping silently. The supervisor I added to catch silent failure was itself failing
silently, and nothing would have said so until an audio server died at a venue and
did not come back. All six now 100755 on disk and in the index; watchdog restarted
and confirmed logging.

Same family as the day-one finding: a thing verified once, in a context that no
longer holds. The new twist is that "chmod +x" is not a fix — recording it in git is.

Also honest-ing the recovery toast (#122). PLN reported needing a ctrl+enter in
Pulsar to resume after a recovery. Likely ICMP port-unreachable poisoning Tidal's
connected UDP socket (ECONNREFUSED on next send, stream stops until re-evaluated).
Until #122 tests it, the toast says "if the music does not resume on its own,
ctrl+enter in Pulsar" rather than claiming a full recovery — a rig that is alive and
silent is indistinguishable on stage from the original fault.

Validation: 16/16 green on the live rig (9 cold + 7 live). Converge no-op verified on
a healthy rig; converge ACTION verified by stopping the watchdog and watching it come
back, with scsynth untouched throughout.

gig-up proved the set COMPILES and said nothing about whether anything could make
a sound. scsynth was read only to colour a banner (:106,:115) and to skip --audio
(:202) — it never failed. This morning the audio server had been dead for two days
and the gate would still have printed GO.

--live is a MODE, not the default. Cold-by-default remains right: cold checks are
safe mid-set and must work on a train. Severity is chosen per link, deliberately:

  HARD  scsynth alive / sclang :57120 / SC -> Ardour (>=24 pw links)
        no audio server, no listener, or no path to the DAW = there is no gig.
  SOFT  Tidal :6010, LCXL -> SC, SuperDirt warmed, perf mode
        all legitimately absent ten minutes before doors; a gate that reddens for
        them is a gate PLN stops reading.

Every check reads the thing that MAKES SOUND — a process, a socket, a real pw-link —
never `systemctl is-active`. That distinction is the whole reason this exists: the
unit reported active/running for two days while the rig was mute, because MainPID is
sclang and scsynth is its child.

ALSO, and it came from PLN's ears: "preload covers set", a HARD cold check.

He heard crackles moving do_it_right -> take_5_drops. Measured in that window: 0
SuperDirt lates, 0 pipewire xruns, 17 lazy soundfile reads — all take5:*. Disk I/O,
not CPU. The preload MECHANISM was healthy and said so ("47/47 banks OK in 1.9 s");
it was warming the wrong list. preload.scd was dated Jul 28, header "10 track(s)
scanned", while the set is 13 — because setlist_samples.py reads setlist_opal2026.txt
(10 tracks) and gig-up/cheat-sheet trust set-coherence.setlist_tracks() (13). The
missing three are do_it_right, take_5_drops, electric_hammer: exactly the pair he was
transitioning between. #120 tracks which source becomes canonical — not guessed here.

Two properties made it invisible, and they are the transferable lesson: preload.scd
is GITIGNORED (correctly — generated), so no diff ever showed it drifting, and
NOTHING COMPARED IT TO THE SET. A generated artifact with no freshness check is a
stale artifact waiting for the worst possible moment; its natural debut was a cold
boot at the venue, on the first play of every track added since Jul 28. It only
surfaced at a desk because SuperDirt had been up since Jul 28 with those banks cached
from a first play weeks earlier, and a restart cleared the cache.

tools/check-preload.sh compares BANK SETS (not bytes — the generator stamps a header
and orders banks, so a byte diff would cry wolf), borrows set-coherence's parser
rather than adding a fourth definition of "the setlist", and has --fix.

Validation, all of it adversarial:
 * check-preload mutation-tested — restore the Jul-28 plan and it names the 7 banks
   that would crackle: daft fbreak120 hammer popkick realclaps superfreak take5.
   hammer and superfreak are independently confirmed: they appear as lazy loads in
   this morning's journal, before any of this was written.
 * --live on the live rig: 7/7 green after fixing MY OWN check (below).
 * THE GATE BITES: run under a PATH shim whose pgrep reports scsynth absent ->
   NO-GO, exit 1, correct fix hint, real rig never touched. That is precisely the
   state that printed GO this morning.
 * cold run unchanged: still GO, live chain not executed.

One check was wrong before it was right, in the session's recurring shape. "SuperDirt
warmed" first warned about a rig that was perfectly warm, because
`journalctl --user -u parvagues-sc` returns lines from a PREVIOUS instance on this
box — it handed back Jul 30 output while the current instance had already logged
"51/51 banks OK". Now filtered by SyslogIdentifier and keyed to the unit's own
ActiveEnterTimestamp, so it asks "did THIS boot warm?" rather than "was there ever a
PRELOAD line", which any previous boot would satisfy forever.

PLN's own live-session edits, made at the desk on 2026-08-01 while rehearsing the
OPAL set. Committed separately from the session's tooling work so the musical
changes stay legible on their own.

* wap — uncomment octersubsubbus 42 / octerbus 41, both on range 0 1.4 "^53".
  The sub layer is live again. (This is the `octersub` want noted in the
  2026-07-31 set-coherence report.)
* bombe_dj — "[808bd:1,808bd:11]" -> "[808bd:11]". One kick sample instead of two
  stacked, on the 1/4/6/9/11 positions.
* do_it_right — mark d3's gM2/gF1 line with a TODO about buttons and variations.

Minutes after test-sc-watchdog exited, PLN got a STICKY desktop notification:

    "3 restarts in 10 min. Giving up — this needs your eyes.
     journalctl --user -u scwd-test-1269280.service -n 50"

scwd-test-1269280.service was the test's own FAKE transient unit. By the time the
notification was read, it had been torn down — so the suggested command returns
nothing, about a failure that never happened, while the real rig was playing fine.

Two defects in one message. The test could reach a real notification daemon at all;
and the give-up path is deliberately urgency=critical/sticky, which is correct for a
real outage and actively harmful for a fake one — it cannot be dismissed by waiting.

Fix: SCWD_NO_NOTIFY=1, honoured by notify() and set by the harness. A test must never
be able to page a human about something that is not real.

Re-run green: 7 passed, 0 failed, and silent. The log still shows the rate limiter
walking (0 prior) -> (1 prior) -> (2 prior) -> GIVING UP, so suppressing the toast
did not suppress the evidence — the assertion reads the log, not the notification.

Same family as the flap it was written to catch: a supervisory tool whose side
effects escape the scope it was reasoning about.

The rig went silent overnight and NOTHING recovered it. Seven layers watched it
happen. From the journal:

    10:37:36  PM: suspend entry (s2idle)
    10:48:38  systemd-coredump: scsynth terminated abnormally
    10:48:38  sclang: "Server 'localhost' exited with exit code 0."
    11:18:47  Bridge: "scsynth process not found"   x63, to the journal only

scsynth is a CHILD of sclang and systemd's MainPID is sclang, so the unit reported
`active (running)` for two days while there was no audio server at all. A green unit
is not sound.

Two findings behind it:

* `parvagues-sc.service` was DISABLED. Six user units start with the session —
  gig-log, bridge, perf-tray, tidal-ardour-autoroute, lcxl-leds-watch,
  midi-autoconnect — and the only one that produces audio was not among them. It had
  been alive since Jul 28 purely as a manual leftover, which is why its death was
  permanent. Now enabled, and symlinked to the repo like the others so it cannot
  drift from the tracked copy.
* The d8346672 fix (QT_QPA_PLATFORM=offscreen) worked exactly as designed: sclang
  survived the suspend perfectly. Suspend killed scsynth instead. The mitigation had
  been applied one layer too shallow — and because sclang stayed up, the unit never
  failed, so no Restart= setting of any kind would have fired.

So: Restart=on-failure (bounded by StartLimitBurst=3/5min, honouring the unit's own
"audio gear should never flap" comment) for the sclang case, plus an external
watchdog for the scsynth case. Deliberately outside the boot path — a bug in
start_and_midi.scd is gig-fatal, a bug in a watchdog is merely annoying. Folding
this into an sclang ServerQuit handler is the right long-term shape; not seven days
before a gig.

THE FIRST VERSION FLAPPED ON PLN'S DESK, and the bug is the interesting part.
The unit had `PartOf=parvagues-sc.service`, so that stopping SuperDirt would stop
the watchdog. But PartOf propagates RESTARTS. Three lines, 90ms apart:

    11:35:46.265  sc-watchdog[1102323]: scsynth GONE — restarting
    11:35:46.321  systemd: parvagues-sc.service: Consumed 1min CPU time
    11:35:46.354  sc-watchdog[1123545]: watching scsynth        <- NEW PID

The watchdog killed itself issuing the restart, Restart=always revived it, and its
in-memory rate-limit counter reset to zero. MAX_RESTARTS was unreachable BY
CONSTRUCTION: the state enforcing the limit was destroyed by the action it was
limiting. It would not have stopped at four notifications; it would have gone
forever. A supervisor must outlive its supervisee.

Fixed three ways, defence in depth:
  - no PartOf on the watchdog unit (the actual fix)
  - restart log persisted to $XDG_RUNTIME_DIR, so being killed cannot launder it
  - await the server's return (up to 100s > TimeoutStopSec=90s) instead of sleeping
    a guessed 25s and judging a boot that was still in flight

Also: every notification was urgency=critical, which Plasma treats as never-expire,
so recovery buried the screen in sticky popups. Stickiness is now the signal —
normal+timeout for "handled itself", critical+sticky only for "this needs your
hands".

Validation. tools/tests/test-sc-watchdog.sh runs the real script against a FAKE unit
holding a FAKE child server, reproducing the green-unit-no-sound topology exactly.
7/7 pass in ~90s and scsynth's PID is unchanged across the run — it touches no real
audio. It pins the flap directly: "watchdog survived issuing the restart", and the
log shows (0 prior) -> (1 prior) -> (2 prior) -> GIVING UP. Plus a 25s --dry-run
against the live unit: zero false positives on a healthy rig.

That testability is the real lesson. v1 could only be exercised by killing the
actual audio server five days before a gig, so its flap was found by a human hearing
four notifications instead of by a test. UNIT and PROC are now injectable.

The watchdog is committed but NOT enabled: it needs one real-rig recovery test
before it earns a place in the boot set.

Two entries. #44's lesson is the one worth mining later: wiring six working
checks into one caller found two of them dead, because every check had been
verified once, interactively, by a human who supplied the missing piece without
noticing. A human types `zsh tools/foo.sh` out of habit and never learns the
file is not executable. An orchestrator has no habits.

#114 records that the overnight constraint list did real work rather than being
ceremony: "never write to a setlist .tidal" is what kept three ranked-"needed"
tasks off the queue, and drawing that boundary first is why the night was
productive instead of risky.

The sweep landed after the brief was written. Replaces the placeholder with the
measured figure and, more importantly, with the right DENOMINATOR: study/ is
documentation prose, so 79+30 of 799 overstates it the same way the harness bug
did. Also surfaces the CRASH class to him directly, since those are the ones
worth his time and the ones I deliberately did not touch.

The clean re-run after the harness fixes. Three numbers have now been attached
to this question and the first two were both about the tool:

    #112 said        14   (a baseline taken during the mute migration)
    pre-fix run     128   (the column-0 `$` artifact, ~49 of them fake)
    clean run        79 BROKEN + 30 CRASH, out of 799 files

And 799 is the wrong denominator too. Split by scope:

    OPAL setlist        13 files    0 BROKEN   0 CRASH
    live/              694 files   51 BROKEN  21 CRASH   ~10%
    study/              59 files   13 BROKEN   3 CRASH
    sandbox/blocks/test 21 files    5 BROKEN   3 CRASH

`study/` is Tidal DOCUMENTATION with inline examples — prose that was never
meant to compile as a track. Counting it inflates the figure the same way the
harness bug did, just more quietly. The number worth quoting is 51+21 of 694.

THE GIG IS UNAFFECTED. Zero setlist tracks in either category, and gig-up is GO.

The CRASH class is new and it is the valuable one. These tracks COMPILE, then
throw when queried — invalid mini-notation inside a string, which is exactly why
the type checker waved them through:

    "[d d] <d d d*2 d*4>]"     unbalanced bracket
    "~ ulgab?:1"               `?` before `:`
    "[0 .. 7 0]"               a range with a trailing element
    "<0.75 .. 0.8 0.8 .. 0.65>"  a sequence on the left of `..`

Every one would fail live, because the scheduler queries every cycle. Until this
run they were all reported as "FAIL — 0 orbit(s) SILENT".

Deliberately NOT fixed here. None are in the set, and each one needs a decision
about what PLN meant — "[0 .. 7 0]" has at least two readings and only he knows
which. Triage is the deliverable; the edits are his call.

Also records a counting caveat rather than hiding it: silent-eval prints a
track's STEM, not its path, so stems that exist in several directories are
marked as ambiguous instead of guessed. The real fix is repo-relative paths in
silent-eval's output — noted on #112, not silently papered over.

Leads with 'nothing needs your hands', because the previous two mornings led
with a fader he had to raise and that framing sets the whole read. Then the two
things that shipped, then the one finding that changes something he was already
told: #112's '14 tracks do not compile' was a harness artifact, and the harness
was calling his own dominant writing style a parse error.

Deliberately does NOT quote a corpus-wide count. The clean re-run was still
going; the number lands on the task instead. Quoting an unmeasured figure is
exactly how #112 got a wrong one.

The previous commit fixed three bugs in the tool whose entire job is catching
failures that make no noise. That makes its own failure mode the worst one
available: saying OK about a track that does not work, or BROKEN about a track
that does. It was doing BOTH, corpus-wide — and it stayed hidden because the 13
setlist tracks, the only ones anyone ever pointed it at, happened to dodge all
three.

Seven tests, all structural — no ghc, no rig, 0.06s:

  1. a column-0 `$` is a CONTINUATION (PLN's dominant corpus style), plus a
     guard that indenting does not FLATTEN relative structure below it, plus
     one for the ordinary inline style — which gets a test precisely because
     dodging the bug is what let the bug live.
  2. a non-zero exit with no SILENT line is a CRASH, and its stderr survives
     into the detail, because stderr is the only thing that says why.
  3. the exit code answers "is anything wrong", never "is anything silent".

MUTATION-CHECKED, because green tests on their own prove nothing. Each fix was
individually reverted and the suite re-run:

  remove the continuation indent  -> test_dollar_at_column_zero_is_indented FAILS
  key the return to silence only  -> test_exit_code_is_not_keyed_to_silence_alone FAILS
  both restored                   -> 7 passed

Test 3 is the one that matters. When crashes were split out of the silence
count, the return value was briefly left keyed to `silent_only`, so a corpus
where every single track failed to compile would have printed "OK — every
declared orbit emits events" and exited 0. That is a false green from the gate,
and a false green here is indistinguishable from a working set right up until
the downbeat.

Full suite: 229 passed.

#16 was ranked "bonus, can be 80/20ed". The cheat-sheet IS the 80/20: at OPAL
he is alone in a field with a laptop and a 48-cell control surface. The FOH
chain and the go-bag he can hold in his head. Twelve columns of CC numbers he
cannot.

GENERATED, NOT AUTHORED. A hand-written card is a copy of tools/lcxl_grid.py,
and a copy drifts silently the moment the grid changes — which it did twice
this fortnight. This reads the grid and the setlist directly, so a card that
disagrees with the rig is impossible by construction. `--md` rewrites
docs/GIG-CHEATSHEET.md; regenerate rather than edit.

THE EMERGENCY BOX GOES FIRST
The moment you reach for this page is the moment something is wrong. A card
that opens with a reference table makes you read past your own panic to get to
the answer. So the top of the page is eight failure modes and their remedy,
each one a symptom he has actually hit: LEDs dark but knobs alive (USB OUT
endpoint stall — replug), a track making no sound (it did not compile; one
block, one error), an orbit that will not die (it is the previous track's
pattern), mixer silent while Tidal looks fine (a fader; the desk wins on touch).

TWO LAYOUT BUGS THE FIRST RENDER CAUGHT
  * the title bar was hand-padded and off by one — now ljust'd like every
    other line, so it cannot drift again
  * gF3 and gM3 hold 7 and 8 orbits, which overran a side-by-side two-column
    family layout and silently misaligned the block. Stacked instead.

AND ONE COLUMN DELETED FOR SAYING NOTHING
The set table first carried a "roles" column. Every track in the set carries
kick+perc+bass+lead, so it printed the identical string 13 times and pushed the
page past 80 columns. Replaced with the orbits each track does NOT declare —
which is the fact worth having on stage, because an undeclared orbit is not
silent, it keeps playing whatever the previous track left there. That column is
the transition ghost list (#77), per track, at a glance. It also shows `perfect`
declaring all 12, i.e. the one track in the set that cannot inherit a ghost.

Verified: 0 lines over 80 columns, renders from a clean checkout, and the
family map printed matches lcxl_grid.filter_family/mute_family exactly.