The rewire (tray Launch Gig -> POST /api/rig, arm-then-confirm) is
written and loads clean, but testing it end-to-end killed Ardour:
launchers.launch('ardour') reported 'launched' while Ardour was already
running, a second instance collided with the first over the session
lock, and both exited -- taking all 12 orbit links with them.

A false negative in is_running() wired to a tray button is worse than
the mislabeled button it replaces, so it does not ship until launch()
fails closed. Parked as a patch beside the note rather than committed
(unvalidated) or left loose in the worktree (this checkout is shared and
another session commits into it by pathspec).

Carries the rest of the session's open ends too: the perf.sh redeploy
that is committed but not live, gig-up's four pre-existing NO-GO
blockers, perf-tray's 2.6GB RSS, and why gig-down is blocked on the
Freebox being down.

Captain's log for the gear-failure pass: one zombie lcxl3-driver holding a
stale binding to an unplugged mk3 explained three of PLN's five symptoms,
the fourth was a mislabeled tray button, and the fifth was tooling that
had been running invisibly all along.

Written for mining later — the sharp bit is that name-based identity is
the actual bug, and it caught the investigation as well as the rig: a
subagent grepped aconnect for LCXL3, hit the driver's own phantom virtual
port, and reported the hardware was back.

lcxl3-driver was policy "login", so rig_units.ensure() started it
whenever it was not already active -- and ensure() runs inside
gig-up.sh --converge, which is what the Bridge's RIG UP calls.

With the CLASSIC LCXL plugged that is actively destructive: the v3
driver has Conflicts=lcxl-leds-watch, so a converge would stop the
painter that was correctly lighting the classic board, then fail to bind
(there is no LCXL3 port) and crash-loop at RestartSec=5. That is exactly
what happened between 11:26 and 14:36 on 2026-09-05 -- 86 restarts, the
board dark the whole time -- and today's repair would have been undone
by the next press of GIG UP.

lcxl-leds-watch was already "manual" for the mirror-image reason, stated
in the comment right below this one. The rule generalises: neither
painter may be auto-started blind, because "which painter" is a question
only the plugged hardware can answer. gig-up.sh's leds() is the
generation-aware chooser -- it greps aconnect for LCXL3 and restarts the
right unit -- and it now owns the decision outright.

Verified: `rig_units.py --ensure` (dry) reports "nothing to do" with the
driver inactive and the v2 painter running, instead of queueing a start
that would have gone dark.

Trade-off worth knowing: on a mk3 rig the driver no longer comes up at
login, so faders reach Ardour only after gig-up runs. That is the price
of not guessing the generation, and gig-up is the documented bring-up.

183 Pulsar SIGABRTs between 2026-08-19 and 2026-09-05, 80 of them inside
one 70-minute window, all from this one block.

perf.sh launched Pulsar in "the script's directory",
$(dirname "${BASH_SOURCE[0]}") -- but perf.sh is deployed root-owned as
/usr/local/sbin/perf-audio (perf-audio.sudoers, for scoped NOPASSWD), so
BASH_SOURCE resolved to /usr/local/sbin and Pulsar got /usr/local/sbin as
its project directory. Coredump argv:
  /opt/Pulsar/pulsar --executed-from=/home/pln --no-sandbox /usr/local/sbin

The Bridge's perf watcher (bridge.py:41 -> perf.py:run_watcher ->
reconcile -> set_mode) shells out here every ~30s, and every mode flag
routes through set_priorities, so the retry was unbounded: crash ->
pgrep finds nothing -> next tick tries again. No backoff, no lock, no
exit-status check. And the "✓ Pulsar launched" line proved nothing: the
verification was a 0.5s pgrep presence poll that caught the doomed
Electron process alive during startup and returned before it aborted.

Fixed by removing the launch, not by resolving the directory better.
Setting priorities is this script's job; starting applications is not.
Ardour, two branches up, has always been prioritize-if-present with no
launch branch -- Pulsar is now symmetric with it. Both the tray and the
Bridge have real Pulsar launchers that pass the right project directory
and surface failures. #116 also wants the hot path unable to start
things by accident, and a root script spawning a GUI editor through
`sudo -u $USER env ...` every 30s is the opposite of that.

NOT LIVE YET: /usr/local/sbin/perf-audio is a root-owned copy and still
carries the bug. It needs
  sudo install -m755 -o root -g root perf.sh /usr/local/sbin/perf-audio
which is outside the sudoers whitelist, so PLN has to run it. Dormant
until then -- the block only fires when no Pulsar is running.

PLN, after the first look at it: "the midi mon is beautiful! so lets
streamline, now it should always oppen, with always on top, all desktops
displayed when i switch from desk 1 to 2 3 4" -- and then "would help to
see the chan num: 33 or 53 etc so as i press i can wire".

Always-on-top needed no change: the WindowStaysOnTopHint at :293 is
unconditional and KWin honours it. The other three did.

IDENTITY. The Wayland app_id comes from desktopFileName, not
applicationName, so KWin saw resourceClass "python3" -- verified by
querying workspace.windowList(), not assumed. Two consequences: a window
rule keyed on the app id would have matched EVERY python GUI on the box,
and launchers.focus_window("midimon") could never match, so clicking the
Bridge/tray button on an already-running lens silently did nothing. Set
desktopFileName + ship a .desktop (which also silences the portal's "App
info not found" warning), and point the launcher's focus at "midiviz".

ALL DESKTOPS is a compositor property, not a Qt flag, so it lives in a
KWin rule (~/.config/kwinrulesrc, wmclass=midiviz, desktops= empty with
desktopsrule=2/Force). Declarative on purpose: it applies at map time,
every time, with no post-launch retry racing the window's existence.
Verified on the live window rather than inferred from the key names --
onAllDesktops=true keepAbove=true.

CC NUMBERS, two layers. Per-cell dim decimals are the map you consult
while wiring; they render in their own pass because f_micro is already
current there, and a per-cell setFont would add 96 font switches a frame
to a path this file keeps to a counted number of drawText calls. They
are skipped when a cell is too narrow to hold them beside the value. And
the header now prints the last-touched control as ^53=127 -- decimal,
with the corpus's caret, so it is the literal token to type into a
pattern instead of a hex value to convert mid-set.

Autostart as a systemd --user unit bound to graphical-session.target,
not default.target: it needs WAYLAND_DISPLAY/XDG_RUNTIME_DIR imported
into the user manager first. Registered "login" in rig_units.py.

Paint path exercised by the offscreen selftest: 116 paints, PASS.

The regen moved timecodes by up to 21s (Piment +21.0, REVOLUTION +19.0, Gimme
Acid +16.7). That file is not just website furniture — Slopmotion cuts visuals
to it, so the release videos were pointed a third of a minute off the music on
a third of the record.

The more interesting finding is a parser conflict the regen walked into.
build_gig_tracksjson.py drafts NEW gig pages and never computes the
styleDistribution that LiveEvent.js renders; gig-metadata.py regenerates
existing ones and does. They also disagree outright on REVOLUTION's style
(jazz vs nujazz). Two parsers, one concept — so the board's line that
build_gig_tracksjson 'is now the way to make any gig page' is too strong and
is corrected here rather than left to bite the next regen.

Left explicitly to PLN: whether the OPAL page documents the set as PERFORMED
(15 tracks, Desire included) or as RELEASED (14). Desire has no v4 boundary
because it was cut, so its v3 timecode rides through unverified. A page can
honestly be either thing; only he knows which one he wants a visitor to get.

SuperDirt's boot runs MIDIIn.connectAll (start_and_midi.scd:19), so SC
subscribes to every hardware MIDI input it can see -- the LCXL included.
The surface then fed SC twice: once raw, once through Midi Through. And
MIDIFunc.cc takes `src` and never checks it (start_and_midi.scd:51), so
every CC fired the handler twice and every button noteOn double-fired
into ~lcxlChordCheck. A double noteOn on a toggle is a no-op: that is how
a mute button silently stops working.

lcxl3-driver's prune() handled this for the v3 board, but only while it
runs -- and it was stopped today (stale binding to an unplugged mk3,
Conflicts=-killing the mk1 painter). A hand-run `aconnect -d` fixed it
live and then SC restarted and re-armed it, which is the whole argument
for putting it in the reconcile loop that already owns the wiring.

Rule 3 matches the RAW board on type=kernel, never on name alone:
lcxl3-driver publishes a virtual port called 'ParVagues LCXL3' that
matches the same name candidates, and cutting the driver's translated
feed would kill the path we actually want. Only surface->SuperCollider
links are cut; -> Midi Through and the aseqdump taps that gig-log and the
Pulsar HUD read all survive, verified live.

Tested by arming the bug (aconnect 20:0 130:3, 20:1 130:4) and watching
the loop cut both and settle: lcxl-path.py went from "LCXL -> SC direct"
back to "LCXL -> Midi Through -> SC", logged once, no per-tick spam, and
gig-log kept counting touched controls.

Also logs PLN's ear-feedback on the classic-LCXL play-test (rose_rouge
works great; midiviz landed) to the archivist's notes.

The first version of this amendment said one OPAL track on SoundCloud was
stale. Then I read the account instead of reasoning about it, and it is
fourteen: the entire Aug-12 first cut is up, thirteen of it under junk
permalinks (/opal26-wap, /12-mafia-12, /ceci-nest-pas-une-bombe) that roll
call could never have found, because roll call only looks where it expects a
track to be. Same error the board made before me, one size smaller.

The finding that explains PLN's ear exactly: eight of the fourteen stale
uploads carry cover art and not one of the six correct v4 tracks does. The
wrong cut is the one that looks finished.

Also corrected: 'every other track matches to the second' was never a verdict,
it was a tolerance. SoundCloud reproduces full_duration to the MILLISECOND, so
at the ±1.5s a transcode seems to deserve, the Aug-12 and v4 cuts of /wap
(361.0 vs 361.6) and /perfect (315.0 vs 314.6) are the same track. That is how
six stale uploads passed a duration check in August.

And two coincidences caught before they could become deletions: /saria-demo
(2025) and /au-revoir-lord-toyota (2020) match OPAL to 50ms by chance on a
132-track account. Duration is a strong discriminator and still not provenance.

    the ones with the cover art were the wrong ones.
    six months of a set played once, and the tell
    was never in the sound -- it was in which
    of the twenty ghosts looked ready.

CLAUDE.md pointed at ../../Web/www/next/content/lives/, which has not existed
since www hoisted the app out of next/ in 6e51fff. The correct path is
../../Web/www/content/lives/{year}/{slug}.md.

This is a worse bug than a typo in docs. The instruction it sits inside is
'Mastering != metadata. Never invent gig metadata... Verify there and cite the
path — if a fact isn't in the canonical source, ask.' A reader who follows the
documented path gets a false 'not found' from the one source they were told to
trust, and the honest next step from there — asking, or worse, filling the gap
from an intermediate script — is exactly the failure the rule was written to
prevent. A stale read is how a Write becomes a delete.

The tide-table scripts already had it right (build_catalog_view.py's LIVES
carries a comment naming the hoist commit); only the instruction was behind.

- process-palette.json wires the editor's command palette to the rig's own
  entrypoints (Init MIDI Channels -> init_midi.sh, and the rest). The whole
  'one entrypoint, fewer files' goal is only real if the editor knows where
  the entrypoints are, so this is rig config, not editor preference.
- sketch.properties points the Processing sketch at its main .pde.
- Logo.xcf is the editable source of the ParVagues mark, from 2025-05. The
  armada design system is documented down to its type stack and its reserved
  brand magenta, and the one file you cannot regenerate from any of that is
  the layered original. 1.2 MB to never have to redraw it.

Deliberately still untracked: TODO.md. It is a paused 2026-06-22 perf/thermal
board, and TASKS_DUMP.md has been the board since 2026-08-16. Committing it
would put a second answer to 'what is next' in the tree, which is the shape
that keeps biting us. PLN's call: fold its two live items into TASKS_DUMP and
delete it, or keep it as a scratch file outside git.

Written earlier today by the gear-repair session and left uncommitted, which
for this particular file is the one failure mode that matters: performance_notes
IS the archive, so an uncommitted entry is not a note, it is a memory.

Two signals, both about the rig rather than the audio:

- *"ok just played rose_rouge works great!"* — first confirmation that the
  CLASSIC LCXL (mk1, 1235:0061) is a fully viable performing surface once
  lcxl-leds-watch owns the paint. Worth recording precisely because the rig had
  drifted into treating the mk3 as the only real board; the classic is a
  fallback, not a downgrade.
- *"and the midi mon is beautiful!"* — on first sight of tools/bridge/midiviz.py.
  A taste signal about the rig's own UI, and the follow-up ask (always-open,
  always-on-top, every desktop) says it plainly: midiviz is an instrument of
  the rig, not a debug tool.

CLAUDE.md asks for every reaction captured durably because this corpus feeds
sampling, effect choices and the auto-mastering heuristics. Durably means
committed.

Four kinds of orphan, all real source, none of it generated:

- tools/sweep_state.py — 'the surface needs a sweep', as a fact the rig can
  act on. Companion to the LCXL3 driver thread.
- armada/tide-table/kick_bass.py — the deterministic kick-vs-bass tiebreaker,
  the one split CLAP cannot do reliably because both live low. Exactly the
  'validate by analysis, never infer role from the name' rule in code.
- armada/tide-table/calibration.stem.json — joins calibration.json and
  calibration.whole-mix.json, which were already tracked. A calibration set
  missing one of its three members is a trap for whoever reads the other two.
- install_superdirt.sclang — pins SuperDirt v1.7.2 and recompiles. The boot
  path is exactly where 'simple, non-brittle, one entrypoint' matters.
- live/collab/benoit/tous_ensemble.tidal and live/midi/nova/nujazz/
  nouvelle_couleur.tidal — two scores. nouvelle_couleur carries a '-- url:'
  header naming its backdrop source, per the scene standard.
- visuals/{maze,nightly_repairs,there,there0}.js — four hydra sketches.

Python entrypoints staged 100755 via --chmod, not left to the working tree.

    the drawer is empty now.
    everything in it was something,
    which is why it took three weeks to look.

Ten shell tools plus FORMAT.md and MIX.md, written 2026-07-29 and untracked
since: stem-stats, energy-profile, onset-detect, silence-detect,
transition-detect, mix-meta, mix-render, mix-normalize, mix-master, lufs-check,
and lib/progress.sh. FORMAT.md defines a <project>.mix.json metadata contract
whose tools 'read existing metadata and augment (don't overwrite unless told)'.

Committed with the exec bit set IN THE INDEX, which is the part that bites:
3fdf5b97 had to restore it on 63 scripts, and MIX.md invokes every one of these
as ./stem-stats. A file that is executable in the working tree and 100644 in
git is executable exactly until someone else clones it. Note also that
'git commit -- path' silently drops a staged mode flip (the trap 1667bb02
documented), so this went in without a pathspec after proving nothing else
was staged.

mix/take94's 833 MB stayed out (a5a5b37). The scripts are the artifact; the
take is the input.

⚠️ Flagged, not resolved: this overlaps ../tidal-ears' mastering CLI, which is
the canonical recipe per MASTERING.md. Two toolboxes that both master a set is
the shape feedback_one_parser_per_concept warns about. I am not merging them
blind — committing first so the overlap is visible in one tree instead of
half-invisible in an untracked directory. PLN's call which one is the spine.

    ten small knives in a drawer nobody opened,
    sharp the whole time.

Written 2026-08-14, uncommitted for three weeks: the -inf fader forensics on
Tidal Live.ardour, the corpus triage that found '51 broken tracks are actually
22 — the harness was the biggest bug', and the 192 kHz master-format audit
('probe, never assume'). All three are marked shareable: true, which is the
whole point — these logs are the source material for the posts and videos,
and they were sitting one bad rm away from not existing.

The rule in armada/tasks/README.md is 'write the entry when the task lands,
while it's fresh'. It worked: the writing happened same-day. It was the
committing that didn't, because git status had become unreadable (see a5a5b37).
Documentation discipline is not one habit, it is two, and the second one is
the one that survives a hard drive.

Left for PLN, one line: all three carry 'log: 025' in their frontmatter, which
025-fifteen-intros-are-not-a-set.md already owns. They also use dated filenames
like 2026-08-03-j5-audio-dropout.md rather than the README's NNN-slug.md.
Renumbering is a call about his own archive's spine, not mine — and the dates
are 2026-08-14, which lands them BEFORE 039, so the next free numbers would
read out of order.

    the log was kept. the log was not carried.
    a ship's record that never leaves the desk
    is a memory, not a record.

Three weeks of untracked working tree, and the reason nothing was committed is
that 'untracked' had stopped meaning anything: 957 MB of take audio and scene
mp4s sat in the same status output as three finished captain's logs, so the
whole list read as noise and got skipped as one.

Sorted by what each thing actually is, against the precedent already in this
file and in the index: mix/'s ten scripts are source and go in, mix/take94's
833 MB does not. The .js scene sketches are source, the 124 MB of rendered
mp4 is not. calibration.json and calibration.whole-mix.json are tracked, so
calibration.stem.json belongs with them; bounds_*.json is generated output
whose builder is already tracked, so it does not.

No new policy here — just applying the one written at the top of this file to
a pile that had grown too mixed to read.

    the tide went out three weeks ago
    and left everything on the sand together:
    charts, driftwood, a gigabyte of water.
    sorting is not cleaning. sorting is deciding what was ever yours.

PLN opened the day hearing it: 'our private cuts of OPAL are badly cut
begin/end unaligned'. The board had one theory (a 24s truncation) and it was
wrong. This lens found the real shape: the ENTIRE Aug-12 first cut is still
on SoundCloud — all 15 tracks — living beside the 6 correct v4 ones under a
litter of ad-hoc permalinks (/opal26-wap, /12-mafia-12, /ceci-nest-pas-une-
bombe), and eight of the stale ones are the ones wearing artwork. So the
wrong cut is the one that looks finished. Of course he heard it.

Duration is the discriminator, and it is far sharper than assumed: measured
on this account, SoundCloud's full_duration reproduces the source to the
millisecond (477.600 uploaded, 477.600 reported). Tolerance is 50ms, not the
seconds you would allow for a transcode — a loose tolerance here buys no
safety, it only invents ambiguity between cuts that differ plainly.

Three things I got wrong first, each left in as a guard:
  - filenames are not authoritative. 'Perfect <3' uploads to /perfect-3, so a
    filename-derived slug reported two live tracks as NOT UPLOADED. Titles
    come from the plan now, which is what the uploader actually used.
  - streaming and club masters share every duration, so offering both as
    rival candidates marked all 14 AMBIGUOUS. Identical generations collapse.
  - a duration collision is not provenance. /au-revoir-lord-toyota (2020) and
    /saria-demo (2025) both matched OPAL to 50ms. Anything uploaded before
    the oldest render on disk cannot be a cut of it, and a coincidence must
    never reach a delete manifest — so the date floor is a hard gate and the
    discards are printed, not silently dropped.

The second pass is the one that earned its keep. Roll call only looks where
it expects a track to be, so it can never see a stale cut hiding under a
different slug. Strays scans the whole account. Absence needs proof.

    fifteen ghosts of a set played once,
    each wearing the cover art of the real thing --
    the lens does not listen, it only counts seconds,
    and seconds were enough.

catalog.generated.json was stale since 2026-06-23 (73 tracks / 23 gigs) and
missing opal-festival-2026 + cosmicfest-2026 entirely — it never got rebuilt
after catalog_view.json picked those gigs up on 2026-09-02. Re-ran the
pipeline (build_catalog_view.py then build_catalog.py):

  catalog_view.json:      81 tracks / 25 gigs / 63 takes (unchanged — was
                           already current, only as_of + minor EDA-coverage
                           deltas moved)
  catalog.generated.json: 73→81 tracks, 23→25 gigs, now carries both missing
                           2026 gigs, both validate clean against models.Catalog,
                           and track-id sets match view↔generated exactly
                           (no silent drops either direction)

Root cause of the as_of staleness: build_catalog_view.py hardcoded
`AS_OF = "2026-06-06"` as a literal instead of stamping the actual build
date, so every regeneration since kept reporting the same frozen date no
matter when it ran. Now computed as `date.today().isoformat()`. Checked the
other suspected bug (a next/content/lives hardcoded path) — not present;
LIVES already points at .../www/content/lives without next/ since 6e51fff.

PLN opened the session hearing badly-cut begin/end on the private OPAL
uploads. Measured all three render generations against what the board
recorded on the account: /take-five-drops sits at 6:57 = 417 s, which is
exactly the Aug-12 first cut — four days and two boundary revisions before
the ear-verified v4. The complaint had a mechanical cause.

Ruled out before believing anything: tracks_v4_streaming matches
segments_v4.json to the millisecond on all 14 (CRIME's 2s rest and
REVOLUTION's reverb tail included), the upload plan's titles/files/durations
are aligned and exact, and the ear file's 19:33 mtime is the signoff record
being written, not a later boundary edit. Nothing needs re-cutting; the
approved audio is already on disk. This is a re-upload problem.

Two things stand in the way and one of them is PLN's call: SCW.delete() is
hard-gated to the test namespace so we cannot remove our own stale uploads,
and SoundCloud skips rather than replaces an existing permalink. Also: no
sc_state.json was ever written, so the account is the only record of what
landed — sc pull --artwork recovers ids, permalinks, durations AND the covers
PLN asked to save, read-only, in one pass.

Plus the catalog answer he asked for (generated.json is 2 gigs stale behind
the view, 14 gigs uncatalogued), www's OPAL tracks.json still on segments_v3
where Slopmotion reads it, the CLAUDE.md next/ path that no longer exists,
and two stale-opens git says already shipped.

veridis_quo (PLN, first jam): hats ply gate moved to ^43 (E3, its own column),
bass line up an octave. rose_rouge: comments from the night it re-validated the
rebuilt rig — a sample-kit idea and the bassline's new name.

The footer froze at '?REC 00:10' on the glass during the first real take. No
code-level freeze reproduced; what WAS wrong: the ghost mirror recomputed state
instead of mirroring device writes (so diagnosis lied), and the 2s tick wrote
unconditionally with no dedup. Now: _home_cache records the string actually
sent, the tick writes only on change (repaint ≤2s, revert ≤2s after stale), and
the 20s label-reassert forces a full repaint so dedup can never mask a device-
side reset. Hardware verdict awaits the next recorded take.

Bug 1: Ardour flushes stems in ~12MB bursts every ~3s — a 2s stop-grace read
every gap as stop/start and reset the clock eleven times in one 20s take.
Grace is now 6s, with the measured cadence in the comment.
Bug 2: transport-button feedback IS bit-gated — /set_surface feedback 8→9
(button-status + heartbeat); the 'not bit-gated' reading died on the live rig.
Bug 3: a mute lens is not a dissenting lens — combine_rec only counts OSC when
arm AND roll are concrete, so 'connected but silent' can no longer stamp
conflict:true on ground truth. src labeling now names actual contributors.
Regression tests drive RecPass through the real flush cadence; 169 green.

Third line of the home screen: 'ParVagues' at rest; 'REC mm:ss' while the
gig-log k:'rec' stream says audio is landing on disk (tail-read of the newest
gig jsonl, same style as last_eval); '?REC' when the files and OSC lenses
disagree — a timer the performer must not trust blindly is marked, not hidden.
Stale or absent signal degrades to the plain footer, never crashes a redraw.
Pure decision function rec_line() + tests.

A GUI-launched Ardour has a BARE argv (only gig-up passes the session as an
argument), so argv-matching alone left the files lens blind on the real rig.
Resolution now climbs: argv → /proc/<pid>/cwd holding a *.ardour → open-fd
symlink targets (most-fds wins on ambiguity, logged once). Injectable proc_root
makes each rung testable against a fake /proc; 9 new tests reproduce the exact
live gap (163 green). check-mix.py shares the latent argv-only bug — flagged in
the sync comment, not fixed this pass.

'Recording X had track Y playing at time Z' becomes answerable. Files lens =
ground truth (stems growing in interchange/, stat-only, 2s grace, RecPass state
machine); OSC lens = Ardour's own word (/set_surface feedback=8 handshake to
:3819, listens :3820, /rec_enable_toggle + transport — the manual's real name
for master arm, not /rec_enabled). Records on transitions + every tick while
recording; lens disagreement ships as conflict:true, never smoothed over.
44 new tests (155 total green), stdlib-only, --no-rec escape hatch.

The straggler from 3fdf5b97's sweep: with core.fileMode=false a PATHSPEC commit
reads modes from the working tree (where they're invisible) and skips a staged
mode-only flip — index-only commits are how mode fixes land here.

lcxl-leds-watch (manual policy) leaves the target; lcxl3-driver and perf-tray
join it. Generated by rig_units.py --target — the ~/.config copy is a symlink
into the repo, so the regen IS this diff.

perf-tray + gig-up: default-on 'Cut GPU at gig launch' (kill TuParles, verify dGPU RTD3, name holders if it won't sleep)

The OLED's track clock resets 'a new track is a new lap' — but skipped the
FIRST load, so lap 1 inherited the constructor's zero. Invisible while the
driver restarted with every session; under the persistent unit (log 039) it
counted the idle hours before the first load: 24:27 on a fresh rose_rouge.

The night's lesson, converged: the driver only ran when a session remembered to
start it, then the session's background task got OOM-reaped and the surface sat
untranslated with every check green.

- tools/lcxl3-driver.service: repo-owned unit like its siblings; retry-forever
  hotplug posture; Conflicts=lcxl-leds-watch so exactly one painter owns the
  board; follow-mode (no --track) so paint tracks the loaded file.
- rig_units.py: lcxl3-driver + perf-tray rows (THE list is complete again),
  new 'manual' boot policy so --ensure can never start the v2 painter into a
  Conflicts= kill; --list --boot filter for shell consumers.
- parvagues-gear: moved into the repo; derives its sets from rig_units --list
  instead of being the fourth hardcoded copy; autostart-on now defers to
  --ensure (it used to enable parvagues-sc at login against policy).
- gig-up.sh leds(): detects the plugged surface generation and hands the LCXL3
  to the driver unit; v2 path invoked via python3 (exec-bit-proof).

core.fileMode=false means chmod fixes your tree and nothing else — the 203/EXEC
class (log 032) recurs on every fresh checkout because the index said 644.
git update-index --chmod=+x is the repo-level fix; every shebang'd script under
tools/ + gig-up.sh + perf.sh now ships executable.

The rig is heavily instrumented and nothing ever looked. gig-log has been
sampling freq, per-core cpu, throttle deltas, fan rpm, per-process cpu/RSS and
xruns at 1 Hz for weeks (512k samples in the current file), yet every failure
this session was discovered by feel, mid-session, at the worst moment. This is
the looking.

Every check exists because it silently cost a real session on 2026-09-04/05:

  * Composed an hour at 800 MHz because the Bridge's perf-watch reasserted
    `silent` every 30 s, undoing each fix within half a minute. Cost 2671
    xruns/hour. Every throttle counter read zero — with HWP, epp=power makes the
    CPU *choose* its floor, so the failure has no error message.
      -> check_perf, check_regime_agreement
  * `gluck` refused to load for 90 minutes. The symlink was perfect; sclang had
    been up 6 days and loadSoundFiles scans folder NAMES once, at startup, so a
    bank added later is invisible. Pulsar's "restart SuperDirt" button is a no-op
    here (SuperDirt is a systemd service, not Pulsar's child) and honestly
    reported 142 h uptime while looking fine.
      -> check_superdirt_fresh
  * Samples/crutch was a symlink pointing at ITSELF since Oct 2025. Dead bank,
    no complaint from anything.
      -> check_sample_integrity
  * A new LCXL3 showed default LEDs because the resolver name-matches the mk2
    product string, while the watcher logged "-> 27 controls" and looked healthy.
      -> check_midi_surface

Three design rules, all learned the hard way:

  1. Green is not evidence. `active`, exit 0 and silence all lied during this
     session, so every check reads the LIVE value from sysfs/procfs/the device
     and prints the number it saw. No check trusts a status.
  2. Every FAIL carries its fix as a runnable command. A preflight that reports
     a problem without the remedy just adds a step between you and playing.
  3. It never mutates the rig. Read-only by construction — a preflight that
     writes is one more un-arbitrated writer, the exact bug class it detects.

16 checks. First run found 1 fail + 3 warns, including a genuine xrun rate of
2224/hour with every CPU line green — a problem we had not noticed and which the
clock fix does not explain.

Known over-sensitivity: check_xruns uses the raw total, and two of the top three
sources (input.hydra_in, the built-in Speaker sink) may not be in the live
monitoring path at all. Xruns on an idle sink inflate the count without being
audible; it should weight by nodes actually in the path.

  gig-preflight.py            human report, exit 1 on any FAIL
  gig-preflight.py --quiet    only WARN/FAIL, for gig-up.sh
  gig-preflight.py --json     for the tray/Bridge

A new Launch Control XL 3 sat there showing factory-default LEDs. Not autostart
(lcxl-leds-watch was active), not wiring (aconnect showed the surface connected
to Midi Through): the port resolver matches the product name "Launch Control XL",
and the mk3 enumerates as "LCXL3 1". So _find_hw_port()/_find_seq_port() both
returned None and every frame went into the void — while the parse stage kept
logging "-> 27 controls" and looking perfectly healthy. A failure with no error
message, which is the worst kind and the whole reason gig-preflight now exists.

NAME_RE also matches \blcxl\d*\b. The inline HUI exclusion becomes SKIP_PORT_RE,
which additionally drops the mk3's DAW and "To DIN" ports — the mk2 exposed one
non-LED port, the mk3 exposes three, and sending a frame to any of them is a
silent no-op.

Validated: resolver now returns hw:1,0,0 and seq 20:0; it returned None before.

NOT fixed here — the LED dialect. The device answered a Universal Device Inquiry
with:

  F0 7E 00 06 02  00 20 29  48 01  00 00  01 01 0B 39  F7
                  Novation  family        firmware 1.1.11

and per Novation's programmer's reference the mk3 lights controls in RGB, one
message per control:

  F0 00 20 29 02 15 01 53 <control index> <R> <G> <B> F7

versus this file's mk2 frame: device id 0x11, command 0x78, a User-1 template,
batched (index, value) pairs, and a bicolor 2-bit red x 2-bit green = 16-state
colour model. So mk3 support is a second dialect plus a replacement colour model
(RGB doesn't map onto 16 bicolor states), and the per-control message form means
a full 40-control repaint costs 40 SysEx messages where the mk2 cost one — the
paint loop will need rate limiting. Tracked in SRE TODO.d.

perf-audio deprioritized Pulsar to nice 10 in cool/extreme and nice 5 in
standard, on the theory that the editor should yield to audio. That trade buys
nothing. The entire audio chain runs SCHED_FIFO *above* it — scsynth 90,
sclang 85, PipeWire 95, WirePlumber 95, Ardour 80 — and a realtime policy
preempts every CFS nice level unconditionally. So Pulsar at -5 cannot take a
single cycle from SuperDirt; it only wins against ordinary processes (browsers,
indexers, builds), which is exactly what we want during a composing session.

Now nice -5 in every mode, overridable with PULSAR_NICE=NN (0 = neutral,
positive = the old behaviour). -5 rather than 0 because the renderer is a single
JS thread and its work is bursty: it has to win the CPU on the tick after a
keystroke, then yield. No root needed — @audio's `nice -19` grant in
/etc/security/limits.conf is what makes the same renice work for pln.

Validated: installed to /usr/local/sbin/perf-audio, re-applied via
`bridge.py perf standard`, all Pulsar processes confirmed at nice -5 live.

Reconnu depuis le strudel 'Vardis Quo' de MariaAreadne: Veridis Quo
(Discovery, 2001). Mélodie d'orgue transcrite en cat 16 cycles,
progression Dm-G-Am-F ralentie x2, arpèges FMRhodes, basse Warsaw,
breaks165/jungle_breaks en stack d8. 116 BPM.

Two commits, 4 files, reviewed as a diff against master.

  tempo_timeline.py       tempo per window in a rendered track, to test whether
                          a planned track was actually played. Answered WAP: not
                          played (median 124.5 over segment 1, nothing at 133
                          above the confidence floor), calibrated first on a
                          declared-160 track which returned 161.5.
  gigmeta sidecar         PLN's confirmed set time (~20:30), his verbatim
                          approval of "Le Jardin", and the WAP question moved
                          from open to resolved with both sources named.
  build_gig_tracksjson    resolve the frontmatter-only fields before writing
                          _provenance, so their origin is recorded whether or
                          not --md-out was passed.
  TASKS_DUMP.md           the board after the release session: six resume
                          points, two of which are the kind that look done when
                          they are not.

Two bugs of mine are in the branch rather than tidied out of it: the octave fold
was an infinite loop for any value a sub-octave band cannot represent, and the
per-frame FFT loop was ~100k Python-level calls for a five-minute stem.

The two-set release board from 2026-09-01 is half discharged. CosmicFest-26 went
from "finished on disk with nowhere to go" to a private 15-item release behind a
canonical www gig page; OPAL-26's remaining upload work is now the only release
thread left, and it is unblocked rather than in flight.

Six resume points, each with the exact command and the wanted output, because
two of them are the kind that look done when they are not: the uploader saying
"N uploaded" is not evidence that N tracks exist (two were accepted, verified,
and then vanished), and renaming the mix's permalink without regenerating the
plan would make the next run re-upload 693 MB rather than skip it.

Also recorded: the www branch is committed but unpushed because that repo is on
github and no key was loaded, the CosmicFest plan still carries `_UNSIGNED` so
nothing may be flipped public, and the two levers left on the record are the LOW
end (the real harmony lever, much bigger than the air shelf) and the deferred
club variant. Plus a note that piment_bresilien.tidal is PLN's own edit.

PLN's set plan lists WAP [133] second at CosmicFest and the recording has no WAP
in it. He remembered dropping it — "didnt play WAP i think in the end [...]
direct from Bombe DJ to Something about Drums, its what i chose that day" — and
asked me to check rather than take his word, which is the right instinct: the
rival explanation is that WAP IS in the recording, inside segment 1, and the
boundary pass read two tracks as one. Duration cannot separate those (track 1 is
6.8 min, an ordinary length here for one track and also for two short ones).

Tempo can, because the two candidates differ: Bombe DJ is 124 and WAP is 133,
which is 0.484 s against 0.451 s per beat — 7% apart, well inside what an
inter-onset histogram resolves. So this walks a rendered track in windows and
reports what tempo each window is actually playing, on the demucs DRUMS stem
rather than the master, because the question is about kick timing and a stem is
far cleaner to find onsets in than a stage-mic mix.

Method deliberately copied from tools/tempo-lens.py, which measures a live orbit
and cannot read a file: spectral-flux onsets, then the MODE of the inter-onset
intervals — never the mean, which over a pattern with ghost notes lands between
two real values and reports a tempo nothing is playing.

Calibrated before it was trusted, on There's Something About Drums (declared
160): median 161.5. Then the answer, on segment 1 with --expect 124 --alt 133:
median 124.5 over 26 windows with a steady pulse, and NO window at 133 above the
steadiness floor. One 20 s window reads 132.5 at confidence 0.11, below the
floor and with no neighbour agreeing — WAP as a played track would occupy twelve
to twenty-four consecutive windows, not one. So: not played. PLN's memory and
the audio agree, and the boundary pass did not miss a track.

Two bugs of my own on the way here, both worth the record:

  * the octave fold was an INFINITE LOOP. Folding into a band of [100, 180) by
    doubling-and-halving never terminates for a value like 99: 99 -> 198 (out
    the top) -> 99 -> 198, because no power of two lands it inside a band
    narrower than an octave. Two processes sat at 98% CPU while I assumed the
    machine was merely busy with the upload. The fold is now closed-form
    (`bpm * 2**-floor(log2(bpm/lo))`) into [X, 2X), and the CLI takes
    --octave-from because an octave is the only honest width: tempo from audio
    is determined up to a factor of two and no further.
  * the per-frame FFT loop was ~100k Python-level calls for a five-minute stem.
    One strided view and one batched rfft instead.

Also, since the meta sidecar is where facts live: PLN's confirmed set time
(~20:30), his verbatim approval of "Le Jardin", and the WAP question moved from
open to resolved with both sources named. And build_gig_tracksjson now resolves
the frontmatter-only fields (time, location, CTA, tags) BEFORE writing
`_provenance`, so their origin is recorded in tracks.json whether or not
--md-out was passed. tracks.json is the file that outlives the invocation.

Side finding for PLN's ears, reported rather than acted on: over the last ~80 s
of segment 1 the drums stem reads 161.5 BPM — the tempo of the NEXT track. Two
readings, and I cannot choose between them from here. Either the segue is a long
livecoded blend with Something About Drums layered in early (which is what
"direct" looks like in TidalCycles), or the boundary at 6:46.6 sits later than
the music does. His ear owns that one.