docs(archive): the set-judge, and what it found on first use

Archive entry for the set-judge build. The entry worth reading later is not the
tool but its first run: it was built to collect go/no-go verdicts and what it
actually established is that the OPAL-26 masters are fine and every middle CUT
is wrong — nine boundaries ear-verified in one sitting, with errors of 3 to 21
seconds that two automated detectors had certified clean.

Also records the four silent traps the build walked into (heads are intros; a
green build with dead playback; three verification probes that tested nothing;
two latent bugs surfaced by reusing existing components).

armada/tasks/completed-archive.md

@@ -1522,3 +1522,45 @@ Commits 6ab09ea, 8d30740. 16/16 green on the live rig.
 
 **Deps.** Closes the assert+converge half of the "big red button". Feeds #116, #119.
 Left open: `--audio` integration, and #122 (does Tidal actually resume?).
+
+---
+
+## Set-judge — go/no-go a whole gig in one pass (2026-08-16)
+
+**Description.** PLN: *"make me a tiny spa to see the heads, soundwaves and
+metadata of what dX are runnning 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."* The call was being made in a music player with the
+answers landing in a chat log, so three weeks later nobody could say why a track
+was cut.
+
+**Done.** `armada/tide-table/build_judge_set.py` (spec-driven; a new gig is a
+copy of a JSON file), `armada/ui/judge.html` + `src/judge/`, `JudgeSet`/
+`DecisionSet` in models.py with TS regenerated, `audio-mounts.json` shared by
+vite and serve.py, an 11-assertion Playwright smoke. Commits `9c6a24c`,
+`20b9151`, `73e9d79`. Narrative log: `armada/tasks/025-fifteen-intros-are-not-a-set.md`.
+
+**Learnings.**
+* **It found the thing it was built to find, on first use** — and the answer was
+  that the masters are fine and the CUTS are wrong. Nine boundaries were
+  ear-verified in one sitting; errors of 3, 4, 9, 10 and 21 s that two automated
+  detectors had certified as clean.
+* **A 30-second head is an INTRO.** Heads are bit-exact with each track's first
+  16 s (r=1.0000), so "1–4 orbits per track" was true and useless — these tracks
+  run 7–10 orbits once open. Judging a set on heads judges fifteen intros.
+* **Green build, dead feature.** `tsc -b` and `vite build` both passed while
+  pressing play threw `No audio loaded` — passing wavesurfer `peaks` + `url`
+  makes it treat a track as pre-decoded and never wire up playback.
+* **Three verification probes in a row tested nothing**, each green or red for
+  reasons unrelated to the app (fetching the empty string and getting the page;
+  `querySelector` not piercing wavesurfer's shadow root; matching `"2 judged"`
+  against `"2/15 judged"`). The probe that works asserts what a human sees.
+* Reusing existing components surfaced two latent bugs: `classify_family` filed
+  claps as Melodic, and `mmss()` had padded a spurious zero onto every sub-10 s
+  time since it was written.
+* **Chain playback was the feature PLN asked for once he used it** — a per-track
+  list hides transitions, which for a livecoded set is the main thing to judge.
+
+**Deps.** Unblocks #147/#152 (publish Sunset Forest). Spawned #177 (re-cut),
+#178 (crossfade-blind detection), #180–#182 (the floor problem and HUD v2),
+#183 (REVOLUTION mix).