PLN: "make me a tiny spa to see the heads, soundwaves and metadata of what dX
are running at that moment / what samples/synth they are, so i can go/nogo /
comment each, then download decisions.json. its a reusable tool tbh we do this
flow often."

That flow was being run in a music player with the answers landing in a chat
log, so three weeks later nobody could say why a track was cut. Ear-feedback is
the scarcest input in this pipeline and it was the only one with no artifact.

Built on the existing A/B Judge rather than beside it: OrbitRail now takes
binS + orbits instead of a whole Take, so one rail serves both UIs. Shapes live
in models.py and the TS is regenerated from them, as before.

FIFTEEN 16-SECOND HEADS ARE FIFTEEN INTROS. The first build showed 1-4 orbits
per track and I nearly shipped that as a finding. The heads are bit-exact with
each track's first 16s (r=1.0000 by correlation against the full track), so the
sparse lanes were TRUE and completely misleading — these tracks run 7-10 orbits
once they open. A ship/cut call on an intro is a call on the wrong evidence.
Activity now spans the whole track; the head stays as a running-order scan,
with a banner saying exactly what it is.

PEAKS ARE THE FEATURE, NOT AN OPTIMISATION. Without precomputed peaks
wavesurfer downloads and decodes the entire file before drawing one pixel —
102 MB per track here. Blank screen at the desk, unusable over wifi on a phone,
which is where the mono-compat check happens. But passing `peaks` + `url`
together makes wavesurfer treat the track as pre-decoded and never wire up
playback: waveform draws, transport looks alive, play throws "No audio loaded".
Both `tsc -b` and `vite build` were green through all of that. Fixed by owning
the <audio> element via `media:`.

THREE VERIFICATION PROBES IN A ROW TESTED NOTHING, each green or red for
reasons unrelated to the app:
  1. fetch(document.querySelector('audio').src) asserting 206 — there was no
     <audio> in the light DOM, so it fetched the empty string, got THE PAGE,
     and read a 206 from somewhere else. It passed.
  2. querySelectorAll('li canvas') — wavesurfer renders into a SHADOW ROOT.
     Playwright's selectors pierce it; querySelector inside evaluate does not.
     It failed while the waveform drew perfectly.
  3. A tally assertion matching "2 judged" against text reading "2/15 judged".
The probe that works asserts what PLN can see: the clock advances and the
button flips to Pause. Prefer the assertion a human could make by looking.

Two bugs fell out of reusing existing parts, which is the argument for reusing
existing parts:
  - classify_family filed vec1_claps, drumtraks, realclaps, clubkick and 808bd
    under MELODIC — it matched perc names by exact-match or prefix, and in
    these scores the perc word is a suffix or an underscore token. A rail that
    shows claps as melody is worse than no rail. Widened to suffix + token
    matching; the test pins that cpluck, dropbass and snippet still must NOT
    match, since those collisions are why exact-matching existed.
  - mmss() padded a spurious zero onto every sub-10s time ("0:002.2"), off by
    one since it was written, in the shared helper. Caught by the smoke test
    reading the clock back.

Titles carried markdown into the release path: tracks.json holds
"There's **Something About Drums**" because backlog.md is prose, and those
strings flow decisions.json -> upload metadata. Emphasis stripped; "<3" and the
shouting caps are PLN's and stay.

Spec-driven so the next gig is a copy of a JSON file, not a code change. The
master->stem map is explicit numbers (keeps summing to 4786.370s, the master's
duration to the millisecond) and the builder asserts it rather than trusting
it. audio-mounts.json is read by BOTH vite's dev middleware and serve.py, so a
URL that works in dev works on the phone over LAN.

Validated: 11/11 smoke green in dev AND against serve.py on the built dist;
22 classifier tests; path traversal out of an audio mount returns 404.