RIFF chunk dump proved samples carry NO semantic embedded metadata (only encoder
tags) — the Pulsar browser shows FILENAMES. So harvest the filename: leading pad
index + instrument-token lexicon → fleet family + source hint. Conservative: opaque
names (JUPI, doing_it_right, 808hc's 'HC') stay family=None → fall back to audio.
Detects kit-like folders (≥2 families by name), the 'jazz is a kit' case.
Corpus coverage: 49% folders / 31% files named, 36 kit-like folders.

One forward pass per folder (text tower computed once) instead of per-file —
~Nx fewer forwards; ffmpeg decode on a thread pool. validate prints a live
[i/N] running-accuracy line. source field now reflects method:mode.

Per PLN: richer ontology + PANNs/AudioSet + ensembles for sample grounding.
- sample_ontology.py: 99 fine descriptors across the 12 families ('this is the
  sound of {a reese bass}'); scored per-descriptor then marginalized to family.
  CLAP fine: 58% -> 68% top-1 (coarse super-family 76%) vs the noisy name truth.
- sample_panns.py: PANNs Cnn14 (AudioSet 527) -> conservative label->family map ->
  per-family prob vector. ffmpeg @32k, zero-pad short one-shots (Cnn14 needs >=1s
  of mel frames or conv5 collapses). Weak on electronic one-shots (AudioSet
  'Clapping'=applause, not a drum-machine clap).
- sample_classify.py: --method clap|panns|ensemble, --fine|--coarse. clap_vector()
  exposes the family-prob vector; ensemble = mean of CLAP+PANNs vectors -> argmax.

Scoreboard (vs name-heuristic, itself noisy): clap-coarse 58% | clap-fine 68% |
panns - | ensemble - (head-to-head primed, not yet run). Stubborn residual =
bass<->kick one-shot (spectral decay tiebreaker is the next lever).

Ground sample families by LISTENING, not by name. sample_classify.py runs
laion/clap-htsat-unfused (transformers, torch CPU) over Dirt-Samples one-shots,
scoring each against text prompts for the 12 fleet families; aggregates per folder
(dominant + homogeneity → kits show as mixed). ffmpeg audio I/O, no librosa.
validate/run/one commands; validate measures top-1 vs the name-confident folders.

Finding (validate): 58% top-1 agreement with the name-heuristic at fine 12-way.
KEY: the name 'ground truth' itself is wrong in many disagreements — CLAP correctly
calls 808hc/808mc congas (perc), which the name-classifier mislabeled bass via '808'.
CLAP is near-perfect on vox/break/clear-bass/kick/keys; the genuinely fuzzy zone is
the melodic cluster (synth/lead/keys/pad). Prompt-tuning is whack-a-mole on noisy
truth. Conclusion: trust CLAP coarsely, not at fine 12-way silently.

PLN-flagged chain of labeling errors, traced to the SSOT classifier:
- 'jazz' was matched to BREAK, but jazz is a multisample KIT (jazz:0=kick,
  :1=snare/hat…). A folder name is not a reliable family signal: it may be one
  family, a heterogeneous kit, or a demucs grab named after a SOURCE song
  (wap, take5, the_revolution, xplosive, rample*). classify_sample_family now
  fires ONLY on names that lexically encode an instrument; everything else is
  None (= needs per-sample analysis). No 'kit registry' (that's name-guessing too).
- removed over-reaching genre/source tokens: jazz, dnb, jungle, loop from break;
  drum from perc. This also FIXES jungle_pads (→pad, was break) and
  jungle_vocals (→vox). amen kept (amencutup genuinely is the Amen break).
- tempo: strip Tidal '--' line comments before parsing cps (ton_numero's
  commented-out morpher no longer counts); a track with a live 'cps (range …)'
  is now flagged morph even when it also declares a fixed setcps. morphing=1
  (Septembre 1er, 60→180), was 0.
- report: + stage_tempo_by_year, sources/roadmap, recurrence gig_slugs,
  classified/unclassified coverage (21% of palette uses need analysis, honest).
- tests: classifier refuses kit/source names; jungle_pads→pad guard. 60 green.

PLN review pass:
- tooltip layer: hover any mark for the tracks/examples behind it. cloud dots
  (track + bpm + date), tempo line dots (median + what it means), histogram
  bands, gig/sample columns, family segments (count + % + example members),
  idiom bars (plain-language gloss of each phrase), staple bars, collab dots
- collab story now plots EVERY track as a dot; tracks at the same BPM pile into
  one larger dot (radius ~ count); median drawn as a tick; sample-tell chips get
  lift tooltips. needs tide_eda collab_fingerprint to emit per-track {name,bpm}
- copy: removed cheap 'not X, Y' contrasts + em dashes per impeccable clarify;
  added 'hover anything' affordance hints

Phase-2 infodesign of the tide_eda corpus EDA: a standalone, scroll-driven
data essay (corpus.html) on the Ship's Bridge dark instrument language.

Six curated stories, vanilla SVG (zero npm deps), measured-width responsive:
1. the slow climb — studio vs club tempo, dual median lines + track cloud +
   histogram + felt-vs-written 2x inset; studio<->club lens is first-class
2. 2024 breakout — vocab burst + gig cadence on a shared time axis, magenta
   reserved for the one earned 2024 accent
3. palette — 12 sample families as a proportion bar, fleet colors + glyphs
4. the accent — signature gMask/gMute idioms (f*16 in 62/73)
5. collab fingerprint — per-collab bpm range + distinctive samples (raph fast)
6. set-staples — recurrence, Cafe trilogy highlighted

- tide_eda: add stage_tempo_by_year (gig-date x track-tempo) for the club lens
- build_corpus.py + 'tide.py corpus': inline-bundle to one self-contained file
  for deploy to me.nech.pl/parvagues/viz (dist/ gitignored)
- IntersectionObserver reveal/lazy-draw, reduced-motion + mobile reflow,
  load-fail banner; lens auto-disables on single-lens sections (honest)
- 59 tests green

Phase-2 infodesign brief from /impeccable shape: standalone scrollytelling
'ParVagues, by the numbers', 6 curated stories on the fleet color language,
studio<->club toggle as first-class lens, vanilla SVG no-deps, deploy static
to me.nech.pl/parvagues/viz. Exact eda_report.json field->chart map for a
clean post-compact build.

- ANCHORS support (slug, end_line) to bound bleeders (mephisteuf/opal-2025/
  38c3-toilet) into clean blocks — no more swallowing neighbour setlists.
- clean_track_line: strip leading set-position labels (Intro:/Finale:/Sunset),
  keep 'Label: Name' / drop 'Name: gloss'; recovers the LIVE-Noctambule pattern.
- leet_fold fallback (B00K→book, T01l3ts→toilets, G00D→good) — resolves only
  after a straight miss so genuine digits (sept1) still match first.
- 9 gap gigs recovered (opal-2024 14/14, algolia-fdlm 26/26, 39c3 13/14…).
- 6 mechanical tests (leet, label, gloss, bpm/transition, ascii/comment skip,
  real-backlog coverage guard). 59 pass.

- models.py: add classify_sample_family() SSOT (token-aware, DM-suffix gated,
  strong-contains fallback) + extend SAMPLE_FAMILIES cues → unclassified 89→34
  uses (~93% classified); honest unknowns (armora, 90s_matrix) stay None.
- models.py: STYLE_ALIASES + norm_style() merges nu-jazz/nujazz, breaks/breakbeat,
  chip/chiptune, hip-hop/hip so the style chart is honest.
- tide_eda + tests use the canonical classifier (DRY); +3 mechanical tests
  (token cases, no-overreach incl. the shil-'oh' FP, style norm). 53 pass.
- regen tokens (cues feed match[] arrays).

Reusable data-scientist pass that finds the stories before any viz:
- TRUE tempo parsed from setcps (corner A), vs metadata bpm (C), with A↔C delta
  (most conflicts are exact 2× half-time/double-time notation, not errors)
- studio tempo (git-creation date) vs stage tempo (gig date) — the creep 110→126
- collab fingerprint: bpm profile + lift-distinctive samples per collaborator
  (raph outed as the fast/club one ~138bpm; nova-solo ~117)
- sample pairings: lift-ranked co-occurrence, ad-hoc 'pair <prefix>' query
- cadence (all 37 canonical gigs), sample-family split, signature idioms,
  set-staples, vocabulary growth (Dirt-Samples symlink mtime, 334-pack Jul'24 burst)
Emits eda_report.json (tidy cuts for the viz phase).

The site tracks.json covers only 23/37 gigs; the rest live in backlog.md as
codename-keyed setlist blocks. New mechanical parser cleans the noisy lines
(ASCII-art, emoji, emphasis, inline [bpm]/{transition} notes, commented-out
tracks) and resolves names to canonical .tidal paths via the catalog's own
alias index (DRY). Gig-codename→slug mapping is authored (confirm:-flagged
when uncertain); suspect_bleed flags blocks that swallowed a neighbour list.
6 gigs cleanly recovered (52 track-slots); seam is larger (dozens of setlists).

ontology in models (SSOT) → generated tokens:
- models.py: OntologyTerm + ColorFamily; ROLE/LIFECYCLE/AGREE terms (color+glyph+
  label) and an authored 12-family SAMPLE-WORLD model (hue-anchored). Variation =
  OKLCH lightness ladder + ±8° hue-fan → 8 shades/family = 96 sample slots. Sample
  axis (conceptual) kept separate from the measured-register axis (ROLE_FAMILIES).
- tools/gen_tokens.py: OKLCH→sRGB (no deps) → tokens.css (armada/ui) + tokens.json
  (triangle fetches at runtime, overrides :root). Validated vs DesignTokens. New
  tide.py 'tokens' step. Never hue alone (glyph+label per term).

triangle UX (low cognitive load, clear hierarchy):
- header calm at rest: one title line, one quiet coverage line, and the agreement
  bar is now the interactive sea-state instrument (overview + legend + level filter
  in one) — replaces the duplicated chips+stack+pills.
- faceted full-text search: free terms + scopes (sample:/gig:/phrase:/take:/style:/
  name:), covers the pattern registry, with 'matched-in' hints so results never feel
  mysterious; '/' focuses, clear button, teaching empty state.
- sound chips colored by sample family (shared lexical classifier from tokens),
  glyph-paired; agreement painted from the ontology (unparsed no longer red).
- tests: 51 green (OKLCH math, shade ladder greyscale-distinctness, ontology
  integrity, unparsed≠conflict, DRY tokens contract, classifier examples).

- drawer 'patterns & kin' section: a track's nearest tracks (shared n-gram
  similarity, clickable to jump) + notable phrase chips (↔ shared idiom /
  ⟳ repeated section), click a phrase to filter the table to who else plays it
- Clusters view toggle: connected components of the neighbor graph rendered as
  family cards (members + the shared phrases that bind them) + a distinctive/solo
  roster. 8 families + 54 solos on the real corpus
- pattern_registry.json loaded best-effort alongside catalog_view.json

- triangle viz loads once: .tidal source embedded per-row in catalog_view.json
  (no per-file fetch / relative-path fragility); take audio resolved at build
  time so a player only renders when a proxy exists; data-load failure banner
- usability: full-width search, fixed colgroup widths, ellipsized sound-lists,
  tooltip'd empty states (corner-C/B gap, not a disagreement)
- tide.py serve: serves the tide-table dir load-once, prints the triangle URL
- #56: drop empty/sketch take-types from candidate links (Take63/64 noise)
- #59 pattern_ngrams.py + PatternRegistry: mini-notation phrases as 'things'
  (unique/repeated/shared), track clustering by sound-token n-gram Jaccard
  (df-filtered + min-shingle guard). 73 tracks -> 1325 phrases, 193 shared,
  33 repeated; surfaces the gMask/gMute boolean idioms + euclid figures
- tests: 43 green (new guards: no empty/sketch candidates, source embedded,
  audio paths resolve; full UT+IT for the pattern registry)

tide.py: one command regenerates every downstream artifact in dependency order
(track_recording_map → catalog_view → catalog → ts_types) + `tide.py test`.

Demote catalog.yaml (hand-state, no valid-as-of) → generated artifact:
- catalog.authored.yaml: the ONLY hand facts (license/collab/inspiration/status/
  notes), each with provenance; keyed by .tidal path
- build_catalog.py: catalog_view ⊕ overlay → catalog.generated.json (73 tracks,
  7 authored), validated against models.Catalog on emit
- catalog.yaml marked DEPRECATED

Triangle drawer = ground-truth validator (per corner):
- A: scrollable, syntax-highlighted .tidal source (fetched live) + parsed orbits
- B: audio player when a take proxy exists
- C: per-gig bpm/style/dur + external gig link + raw ingredient list (site's claim)
- search now spans ingredients/samples too
- enrich catalog_view rows with raw ingredients (+ Ingredient/Catalog models, TS)

Serve from the REPO ROOT so source/audio resolve:
  python3 armada/serve.py --dir . --port 8731
  → /armada/tide-table/triangle.html

30 pytest green.

Reconcile every track across three corners and light the cells where they
agree or fight, so "is our analysis good?" becomes a measured number.

- build_catalog_view.py: reconciler over 73 tracks; pure build() + thin main(),
  validates against the pydantic CatalogView on emit
- tidal_score.py: parse indented do-block dN (recovered burn_this_book + 13 more)
- agree() taxonomy: unparsed / no-claim / agree / partial / conflict / divergent
  — a parser miss is NEVER reported as a disagreement (the cardinal fix)
- models.py CatalogView → gen_ts_types.py → types.gen.ts (DRY single source)
- tests/: 28 pytest — parser regressions, agree taxonomy, IT-on-real-data with
  coverage regression guards, DRY contract (mechanically tests the metadata prior)
- triangle.html: Ship's Bridge lit-cell dashboard (served by serve.py)
- tasks/013: captain's log

Result: 47 agree / 5 partial / 2 conflict / 3 divergent / 16 no-claim;
recorded 40/73; EDA coverage 1/63 (the gap, now visible). The first conflicts
flagged were our own parser bugs — metadata was righter than the machine.

Reframe (per PLN): we split the MASTERED set into shippable units (tracks / EP /
gapless BC album); we don't touch the gig recording. Source is swappable (online
streaming_final today; a re-rendered master can drop in).

- align: measure proxy→master offset by xcorr of energy envelopes (proxy from
  stemmap, master decoded). Coarse 2s + fine 0.25s + two-half drift check.
  Result: master_t = proxy_t + 2.75s, stable across halves (peak 0.69-0.80) —
  confirms the manifest's calibrated +3. (Caught + fixed a sign bug on first pass:
  measure, then verify your own number.) Per-source: the +3 was the gig-GT, this
  is the stem-render; same ballpark, but measured not assumed.
- frag: audition cut variants for a boundary from the real master — context,
  standalone clean edges (trim bleed + fade), and xf_direct (WAP tail acrossfades
  into 'Plosive head, the bordel dropped = PLN's 'crossfade direct de 0326 a 0345').
  Skip window auto-set from the v2 bleed detector. Each render self-verified non-silent.
- WAP->'Plosive (cut4) rendered: 5 frags @ /frags/ for phone audition.

Next: split (hybrid gapless-exact + standalone-fade outputs) once PLN picks edges.

v1 only saw a foreign ORBIT lighting up on the wrong side of a cut; it was blind
to a foreign SOUND on an orbit active on BOTH sides (e.g. breaks playing through
a transition). v2 (--spectral) closes that:
- per shared orbit, profile tail/head probe windows + each track's interior ref,
  composite timbral distance (band-shape cosine + centroid shift); flag when the
  tail sounds like the NEXT track (incoming) or the head like the PREVIOUS (outgoing).
- gated by MIN_REF_DIST (versions must differ) and SPEC_MIN_RMS=-50 (probe must
  have real sound — the silence gate caught 4 of my own false positives).
- independently flagged the WAP->'Plosive breaks bleed (orbit-08, Delta=0.475) that
  v1 missed = PLN's 'bordel jusqu'a 0342' zone; + Premier Septembre->Techno Orage.
- EDL now 9 ear + 7 orbit + 6 spectral derived rows; round-trips through MasterEDL.

Delta = severity rank; live crossfades blend by design, so PLN's ear gates feel.

#40 Judge annotations were stale+wrong: labels came from a hardcoded ROLE_MAP
(orbit-03 'hats' though punkachien.tidal plays 'dr'; d6/meth_bass missing) and a
hard -38dB on/off gate hid audible orbits that dip within a 2s bin (PLN saw only
d3 OR d8 while hearing both).
- tidal_score.py: parse per-orbit sound map from any .tidal (reuses tfidf vocab;
  also accepts the bare $ "sample" source form). dr/meth_bass/jungle_breaks now correct.
- audio_lens: classify_family() (breaks->tops, drums->percs by identity, register
  by measured centroid), profile peak_db. Fix orbit_files to try padded+unpadded
  channel names (Take89 uses 'Tidal 01-1' -> centroids were silently None).
- build_player_data: labels from score, family validated by centroid, two
  thresholds (litFloorDb -52 visible / activeDb -38 driving), emits validated PlayerData.
- OrbitRail: graded activation (dim->vivid), no orbit vanishes when quiet.

#42 DRY: pydantic models in models.py are the single source of truth; gen_ts_types.py
generates ui/src/types.gen.ts (types.ts is now a re-export). No more hand-synced shapes.

#41 WaveformPlayer: Audacity-style zoom (+/-/ctrl-wheel) + scroll, drag-select a span
and loop it (Regions plugin), keyboard L=loop. tsc + vite build green.

Embodies the division-of-labor principle: the machine flags OBJECTIVE boundary
leaks so PLN's ears judge only feel. Detector = exclusive-orbit ignition: at each
cut, an orbit core to one neighbour but not the other, lit on the wrong side (down
to a sensitive -55dB floor for quiet foreign onsets).

Validated on Take89 (Montreuil): independently caught BOTH of PLN's hand-flagged
bad_cuts — cut1 (Take5 bass pre-echo, "changement de basse avant le cut") and cut5
(Plosive→Jeudi Drill, the Bogdan stabs on d9/orbit-09, ~-48dB) — plus 5 new
candidates. Auto-populates master_edl_take89.json: 16 edits = 9 ear + 7 derived.

Mapping nailed down (was wrong first pass): stemmap orbit-NN == dN; Tidal's
`# orbit X` is SuperDirt-0-indexed → d(X+1) (so `# orbit 8` == d9). d8 = always
breaks. Honest coverage cap logged: a foreign SOUND on a shared+active orbit is
invisible at orbit level → v2 = spectral/per-sound novelty.

PLN: "samples and synths are the same class for fingerprinting." sample_tfidf now
indexes synthdef names (SC synthdefs/ + SCLOrkSynths quark + SuperDirt builtins)
alongside Dirt-Samples, tags each sound sample|synth, and persists kinds. vocab
871 sounds (293 samples + 63 synths used). septieme_armee signature now surfaces
moogBass/FMRhodes1/bassWarsaw — but all common (no rare tell): its identity is the
orbit-arrangement (the SNA riff), not a signature sound. L3 needs both signals.

Take35 identified by blind ear-test as Septième Armée (Seven Nation Army cover,
septieme_armee.tidal, 4:35), NOT the 38C3 "Pitbul Punk" the ±3d date-join guessed.
Corroborated by orbit→sound map (d4 bassWarsaw = the riff bass, etc.). take_gig_map
corrected; performance_notes logs the find + cover-license caveat.

audio_lens.py consolidates the spectral toolbox (profile / stemmap / mix) so the
machine catches OBJECTIVE audio defects autonomously, reserving PLN's ears for
subjective feel.

- profile(file): band-energy % + centroid → is_broadband() gate (centroid >140Hz
  and >18% energy above 150Hz). A pure-kick mix is centroid ~111Hz, mid<12% → FAIL.
- mix(take, --orbits): EXPLICIT numpy sum of local interchange stems (no amix black
  box, no freebox), peak-normalized, then SELF-VERIFIES the render is broadband and
  exits non-zero on FAIL. Caught + fixed the kick-clone bug that shipped twice.
- stemmap(take): per-orbit activity (driving/sparse/silent) for arrangement reads.

Take35 mix rebuilt via the tool (v3): cen 168Hz, mid present, PASS — replaces the
two broken kick-clone mixes (deleted). Lesson: measure your own output; extend the
toolbox, no /tmp throwaways.

EDL: seed_edl_take89.py + master_edl_take89.json (9 typed edits, 3 bad_cut labels)
from PLN's WH-1000XM5 review, logged in performance_notes.md. tasks/011 captures the
TF-IDF + verify-your-output story. #25 reframed: auto-detect boundary sample-leak.

First versioned import of L'Armada (media/marketing toolbox) plus this session's
tidalcycles→tracks tooling work. Audio kept out of git (armada/.gitignore).

Tooling:
- tools/sample_tfidf.py — sample TF-IDF over the local .tidal corpus (772 docs).
  Vocabulary = local Dirt-Samples folder names (custom packs symlinked in);
  sound-context extraction (s/sound/#) so mininotation booleans don't pollute.
  Derives discriminative samples: PunkAChien = vec1_acid(df18)+cpluck(df23);
  jungle_breaks(df78) is common filler. → sample_tfidf.json (validated).
- armada/tide-table/models.py — shared pydantic v2 models: Provenance, LocateCell,
  MasterEDL/MasterEdit, TfidfReport. Typed, JSON in/out, provenance-carrying.
- armada/tide-table/locate-matrix.md + build_track_recording_map.py — L0 metadata
  map (site tracklists × take_gig_map), canonical track id = .tidal path.
- armada/tide-table/seed_edl_take89.py + master_edl_take89.json — PLN's phone+
  WH-1000XM5 resplit review as 9 typed edits (3 bad_cut = detector ground-truth).
- armada/serve.py — Range-capable LAN static server for phone audition.
- armada/ui/ — Vite+React+TS+Tailwind Judge pilot (Ship's Bridge design system);
  PRODUCT.md + DESIGN.md design language.

Correction captured: the "Hamburg PunkAChien" was a misID (Insouciance→Liquid
Finale @ 39C3); real 38C3 PunkAChien = Take35 (pending ear-verify). meth_bass/d6
silence is a controller-ergonomics artifact, not a track tell — dropped as a
fingerprint feature. See performance_notes.md, tasks/010 + 011.

Sync-up of the workspace: edits across nova/collab tracks
(quand_on_decolle, take_5_drops, punkachien, jeudrill,
sunny_side_up, plosive, wap, perfect, techno_orage, ...),
new collab/solo pieces (mains_libres, the_revolution_will_be_sampled,
bois_bumbum, do_it_right, bullet_train), backlog/CLAUDE.md
project notes, perf.sh (CPU governor helper), and the
analyze_samples / fan_check / shipowiz tools.

Pulsar editor plugin rendering a track-aware HUD topbar
(role colors, LCXL spatial layout, header_comments). Was
sitting loose at ~/Work/Tools/pulsar-parvagues-hud with no
git; vendored into tools/ so the package travels with the
livecoding workspace and can be shared in one clone.

node_modules excluded; run `apm install` (or `npm install`)
inside the dir to set up.

28 output channels (+4), 14th orbit dedicated to tidal-ears audition.
No Ardour track needed — captured via PipeWire virtual sink.

TODO verify together: bass rewritten from pyin analysis (librosa),
kick LPF raised for click, d3 cut groups for OH ring-over, cello updated.
Bass was wrong — NTO uses sustained Bb1/F2 alternation with chromatic
slides, NOT F/A/C arpeggio. A2 never appears in any segment.

d5: keep dope slow-2x base, ^58 now layers fast 2 (original NTO speed)
d9: 90s_synatm:8 (99.7% low, 0% hi — warm, not windy)

d4 bass: octersubbus replaces lpfbus on ^53 (user preference)
d5 arp: 16-bar cycle with A3/C4 variations (demucs melody arc)
  Bb/F anchor bars + A3/F4 + A3/C4 + Eb/Gb variations
d7 cello: tighter room (0.25), shorter legato (1.2), rhythmic
  answer phrase on beat 3, mirrors bass V chord one octave up

Every voice now has subtle 32-bar sweeps matching NTO's arrangement evolution:
- d1 kick: lpf 700→250 (darker), room 0→0.15 (opening)
- d2 clap: gain 0.5→0.8 (grows into mix)
- d3 OH: n 5→9 (brightening), hpf 2k→5k, gain 1.2→1.5
- d4 bass: room 0.05→0.2 (space opens)
- d5 arp: room 0.2→0.5, sz 0.3→0.7 (tail grows)
- d6 piano: modIndex 1→3.5 (richer harmonics)
- d7 cello: room 0.3→0.6, gain 0.8→1.1 (presence grows)
- d9 pad: lpf 1.2k→6k (warm→bright), gain 0.7→1.0 (swells)
All manual controls preserved — automations are background evolution.

Full NTO drum character applied:
- d1 kick: att 0.05 (round thud), lpf darkening over time (saw 800→300)
- d2 clap: quiet (gain 0.7), dry (room 0.1), hpf 800 — mid-snap only
- d3 OH: evolving brightness (n6→n8), att 0.04 (soft onset),
  ^44 toggles whisper-shimmer mode (gain 0.3, hpf 6000, legato 0.15)
- Header: full drum voice analysis with centroid evolution,
  attack/decay times, peak levels per section

Deep demucs drum study findings:
- Kick: 8th note pulse (NOT 4otf), ghost doubles before drop
- Cymbal: THE defining texture, upbeat 8ths throughout most of track
- Clap: appears late (~200s) on contretemps (beats 1&3), ~1200Hz dry snap
- Drop at 2:20: kick disappears, pure cymbal upbeat 8ths
- Arrangement: kick 8ths → cymbal break → kick+cymbal+clap climax

Changes:
- d1 kick: rewritten as 8th note pulse, ^29 ghost double control
- d2 clap: ^43 ON = contretemps (beats 1&3), OFF = backbeat
- d3 hats: OH on every upbeat 8th, ^44 toggles cymbal breakdown mode

Major findings from demucs onset analysis:
- NTO bass uses ONLY F2/A2/C2 (V chord tones) — never sits on Bb root!
- Bass is rhythmically alive, not held notes. Varies each 2-bar phrase.
- Guitar(arp) stem evolves: F3/F4 base → Bb3/A3 mid → Gb4/E4 breakdown
- Piano: sparse Bb3 stabs with F3/A3, only in specific sections

Changes:
- d4 bass: rewritten as rhythmic V chord figure (f2/a2/c2 variation)
- d6 piano: NEW — sparse FMRhodes1 stabs, ^31 knob controls density
- d7 cello: reworked as counter-melody (sustained 5ths/3rds)
- d9 pad: fixed cut group collision (was cut 7, now cut 9)

NTO La Clé des Champs algoraoke:
- Bass: dropped to Bb1/F2 (demucs confirmed 58-87Hz range)
- Bass answer: V chord tones A2→C2→F2 (onset analysis, not Eb Db C)
- Answer timing: quarter notes over 2 bars (was 83ms triplets, now 500ms)
- Arp: simplified to <[bf4 f4]*8!6 [ef4 gf4]*8!2> (no note-flip)
- Cello d7: reworked as sustained counter-melody (5ths/3rds), not bass double
- Drums: added sometimesBy knob control (^30 clap flam, ^32 hat roll, ^52 legato)
- d8 breaks: full treatment (^92 ply, ^60 chop mode, ^36/^56 break switch)
- Header: rewritten with demucs spectral evidence and onset analysis data
- SC: s.volume=2 + StageLimiter.activate for output levels

Research + MIDI analysis of the original track:
- Key: Bb minor (harmonic minor dominant F)
- Arp from MIDI: 2-note 16th oscillation (Bf4/F4, order flips per chord)
- 16-bar cycle: Bbm(4) | F(4) | Bbm(4) | Ebm(4)
- Drums: 808bd+clubkick, h2o handclap, h2o hats (sparse offbeat)
- All melodic voices on MiOmi (bass, arp, pad, lead)
- Timbral faders for live "wider before drop" sweep (FM/detune/cutoff)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>