docs(tasks): archive #177 and #178 — the re-cut, and why the detector was closed

Written as documentary source material: three detectors' numbers, why the
best-looking one was the least trustworthy, and the reframing (onset vs
takeover) that a failed lens produced.

armada/tasks/completed-archive.md

@@ -1564,3 +1564,67 @@ vite and serve.py, an 11-assertion Playwright smoke. Commits `9c6a24c`,
 **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).
+
+
+---
+
+## #177 — Re-cut OPAL-26 v4 from ear-verified boundaries
+
+**Description.** The v3 cut of "Sunset Forest" (OPAL Festival, 2026-08-08) had wrong track
+boundaries. Blocked the whole release.
+
+**Done.** Every boundary settled and the record rendered.
+`dab8d38` `d11e6c9` `5f7db02` cut-lens v3 (dead end, see below) · `d3dbe59` #1's head trim
+measured · `192c6e9` the boundary lab · `af60c81` four ear calls merged · `6099df8`
+`apply_boundaries.py` + segments_v4 · `30e62f4` the reusable pipeline · `7b9dd17` the seam
+edits · `10829fd` subtle reverb · `db6529d` the apad fix.
+
+Result: 14 tracks × streaming + club, durations exact, 48 kHz/24-bit, plus drop-free
+continuous mixes at 4358.94 s. Desire cut; #13 Vague de CRIME is the last official track.
+
+**Learnings.**
+
+*Automated cut detection is a dead end for crossfaded sets, and it took three attempts to
+earn that conclusion.* v1 missed by 24 s median, v2 by 16.6 s, v3 got to 3.7 s — but on n=2.
+Then the "better" gate HALVED coverage (8/14 → 4/14). The real result was never an error
+number: **#13 and #15 produced zero voting orbits across three gates and two estimators**,
+because those transitions share their sounds and a timbral lens is blind there by
+construction. The seductive part was v3 run 2's 0.4 s median — computed on the two boundaries
+that already had ear answers. A metric on the cases that were already easy measures nothing.
+
+*The failure was still worth the effort, because it explained itself.* Every v2 error was
+NEGATIVE, and the accurate ones were the boundaries PLN called short: the lens finds where the
+incoming track first APPEARS, the ear marks where it TAKES OVER. That reframing produced the
+span model and, ultimately, the tool that actually worked — offer candidates, let the ear pick.
+
+*There is no universal fraction-through-the-crossfade.* PLN's calls on two ADJACENT cuts
+disagreed: `takeover` on #11 (machine within 0.13 s), `mid` on #12 (machine 8 s late).
+
+*A track cut from the release still owns its boundary.* Desire was dropped, so the edge INTO
+Desire looked like dead work — it is where the last official track ENDS. Nearly missed.
+
+*A set is released twice.* Cutting Desire fixed the tracks; both continuous masters still
+played it in full. PLN caught it. The continuous upload is also the lower-risk rights path.
+
+*Verify against intent, not output.* The apad bug surfaced only as a 0.24 s duration
+mismatch, and the tempting fix (`atrim` back to the prediction) would have gone green while
+clipping the ring. The prediction was wrong, not the audio.
+
+**Deps.** Blocked #179 (uploader), #183 (REVOLUTION mix). Both now unblocked.
+
+## #178 — Crossfade-blind cut detection — CLOSED as a dead end
+
+**Description.** Build a detector that finds track boundaries across a 4-cycle crossfade,
+where activity-based detection is structurally blind.
+
+**Done.** Three lenses built and honestly validated; all recorded in-code with their numbers.
+Declared a dead end in `cut_lens3.py`'s docstring rather than tuned a fourth time.
+
+**Learnings.** Recorded above and in `reference_boundary_detection_three_lenses`. The two ways
+out if ever revisited: change the FEATURE (harmonic/key change, or sample-bank identity from
+the score, which differs even when the spectrum does not), or accept ear-only. Also worth
+fixing first: Montreuil26 has 14 user-confirmed boundaries but NO per-orbit stems, so it
+cannot validate a stem-based lens — that gap should be closed before attempt four.
+
+**Deps.** Superseded in practice by the boundary lab (#B2), which made the ear call cheap
+enough that the detector stopped being necessary.