docs(tasks): log 014 — ParVagues Unwrapped explorable viz

armada/tasks/014-parvagues-unwrapped-the-corpus-you-can-hear.md

+---
+log: 014
+title: "ParVagues Unwrapped — the corpus you can hear"
+date: 2026-06-07
+task: "#81 superfeature shortlist → exploratory viz (under #64)"
+tags: [tooling, dataviz, audio, mda, ux]
+shareable: true
+---
+
+## Cap (what & why)
+The MDA found the story (5 superfeature axes, timbre ≠ family); now make it
+*explorable*. Not another scrollytelling essay (that's `corpus.html`) but an
+instrument: drop all 1,485 samples into feature space, let the eye AND the ear
+roam — pick the axes, recolour by lens, click a dot to hear the sample.
+
+## Manœuvre (how)
+Two layers, parsers-over-copy:
+1. `build_unwrapped.py` bakes one `unwrapped.json` — reuses `feature_eda.load_matrix`
+   so the matrix matches the MDA exactly, then adds the three things a scatter needs
+   per file: the PCA projection (PC1–5), the KMeans timbral-cluster id, and an
+   audition `.wav` path (a gitignored `_samples` symlink → Dirt-Samples). The 5
+   leading PCs name themselves from their loadings, oriented by a stable anchor so
+   the poles always read right. Contract test guards the baking.
+2. `unwrapped.html` — zero-dep, Ship's Bridge dark, fleet family colours. Canvas
+   scatter (axis-pick × 4 colour lenses × search × hover × click-to-audition) +
+   four MDA panels (loading bars, RF importances, per-family beeswarm, family×cluster
+   grid). Magenta reserved for the audition pulse — the *now*.
+
+## Prise (findings / artifacts)
+- `armada/tide-table/unwrapped.{py,json,html}` + `tests/test_build_unwrapped.py` (4 green).
+- 1,485 samples · 36 features · 12 families · intrinsic dim 19 (90%); 5 PCs = 59% var.
+- Verified headless (Playwright): no console errors, audition fires a real WAV request
+  (`_samples/dr_few/005_020.WAV`), all panels populate. Served via `serve.py`, opened in FF.
+
+## Sel (the shareable learning)
+The PCA axes named themselves into *musical* qualities — Brightness, Timbre colour,
+Loudness, **Envelope (kick ↔ bass)**, Tonal↔noisy — without anyone labelling them.
+The kick↔bass axis the resolver struggled with literally falls out as PC4. And the
+family×cluster grid is the receipt for "a folder is a loose grouping": at ARI 0.25 the
+audio clusters refuse to line up with the filename families.
+
+## Hameçon (hook)
+A map of every sound I livecode with — and you can click any dot to hear it. The
+machine drew the axes; turns out it drew *brightness* and *punch*.