docs(archive): A3-b — the OPAL SoundCloud diagnosis, for the long term

Structured archive entry for the session, written as standalone learning rather
than a board line. Fourteen findings, of which the four that will outlive the
task: SoundCloud accumulates generations and never replaces (so a stale upload
is permanent until deleted); the wrong cut wore the cover art, which is why
PLN's ear beat every tool to it; the 'matches to the second' tolerance was the
bug, because full_duration is millisecond-exact and a loose window made two
different cuts identical; and roll call cannot find a stray, because a check
that only looks where it expects a thing never discovers it elsewhere.

Written for a reader months out with no memory of any of it.

armada/tasks/completed-archive.md

@@ -1814,3 +1814,91 @@ suite 112 → 128.
 
 **Deps.** Built on #20's `engine/stempack.py`; produced the data #21 browses. Unblocks #22
 (the acapellas, instrumentals and ~27 demucs catches) with no new code.
+
+## A3-b — The OPAL SoundCloud cuts: a stale generation, not a bad cut (2026-09-05)
+
+**Description.** PLN: *"on soundcloud i see our private cuts of OPAL and they are
+badly cut begin/end unaligbed we need to reupload and save their images tho wheres
+out catalog at?"* Diagnose why the ear-verified v4 release sounds wrong on the
+account, preserve the artwork before touching anything, and report where the
+catalog stands. The board's standing theory was that `/take-five-drops` had
+uploaded "24 s short" while every other track "matched to the second".
+
+**Done.**
+- `master sc pull --out ~/Work/Sound/Prod/sc_pull_2026-09-05 --artwork` — read-only,
+  token still live, no login. 132 tracks, 10 playlists, **101 artwork originals**.
+  Also the state recovery: no `sc_state.json` had ever been written.
+- `armada/tide-table/sc_generation_lens.py` (`fe8ec16`) — roll call + whole-account
+  stray scan, duration-keyed, date-gated, plan-authoritative titles.
+- `armada/tide-table/build_delete_manifest.py` (`bd3a680`) and
+  `master sc delete --manifest` in tidal-ears (`ae1ca0c`, merged `ab3285b`) — the
+  "explicit, separate call" `scwrite.py` had promised since August. Dry-run
+  default, identity verified in both directions, public refused, per-entry reason
+  required, batch capped before any API call, full receipt. **Dry run: 14
+  deletable, 0 refused**, audited independently. Nothing deleted — `--go` is PLN's.
+- Catalog rebuilt (`f1fa484`): generated 73→**81 tracks**, 23→**25 gigs**, now
+  identical to the view, OPAL and CosmicFest present.
+- OPAL `tracks.json` regenerated from v4 (www `5654047`) — was off by up to 21 s.
+- `tools/ardour_session.py` + check-mix fix (`27f2351`); tree triage 39→5 paths
+  (`039ff57`, `71516e2`, `a14aee7`, `f5d3461`, `ec9c7d4`, `1b03cb0`); CLAUDE.md
+  path fix (`577767a`); log 040 + resume points (`33794e3`).
+
+**Learnings.**
+- **SoundCloud accumulates generations.** It will not replace the audio behind a
+  permalink (paid feature) and our uploader is idempotent by permalink, so every
+  re-upload under a fresh slug ADDS a copy. Nothing ever replaces anything. The
+  entire Aug-12 first cut of OPAL-26 was still up — 15 tracks — beside the 6
+  correct v4 ones, under junk slugs (`/opal26-wap`, `/12-mafia-12`,
+  `/ceci-nest-pas-une-bombe`).
+- **The wrong cut wore the cover art.** 8 of the 14 stale uploads had artwork;
+  none of the 6 correct v4 tracks did. The superseded audio looked *finished* and
+  the correct audio looked like drafts. PLN's ear was reading the artwork. A
+  completeness signal attaches to whatever was last fussed over — usually the
+  older thing, because it had more time.
+- **`/take-five-drops` was never "24 s short".** It was 417.000 s — to the
+  millisecond, the Aug-12 cut. Not truncation, a different generation. The board
+  had been carrying a plausible wrong cause for three weeks.
+- **The tolerance was the bug.** SoundCloud's `full_duration` reproduces the
+  source **to the millisecond** (477.600 → 477.600), so the working tolerance is
+  0.05 s, not the seconds a transcode seems to deserve. At ±1.5 s the Aug-12 and
+  v4 cuts of `/wap` (361.0 vs 361.6) and `/perfect` (315.0 vs 314.6) are the same
+  track — exactly how six stale uploads passed an August duration check.
+- **Roll call cannot find a stray.** Checking each *expected* permalink
+  structurally cannot see a stale cut under a different slug; the whole-account
+  scan is what found 13 of the 14. A check that only looks where it expects a
+  thing can never discover that thing somewhere else.
+- **A duration collision is not provenance.** `/au-revoir-lord-toyota` (2020) and
+  `/saria-demo` (2025) matched OPAL cuts to 50 ms by chance on a 132-track
+  account. Strays are now hard-gated on a date floor and the discards are printed.
+  A coincidence must never reach a delete manifest.
+- **Refuse to re-render before measuring.** The obvious move on "the cuts are bad"
+  is to re-cut. All 14 v4 FLACs matched `segments_v4.json` to the millisecond
+  (CRIME's 2 s rest and REVOLUTION's reverb tail included), and the ear file's
+  mtime postdating the renders was the *signoff record* being written, not a late
+  boundary edit — checked specifically because it looked like a smoking gun.
+- **Filenames are not authoritative for a permalink.** `Perfect <3` uploads to
+  `/perfect-3`, so a filename-derived slug reported two live tracks as NOT
+  UPLOADED. Titles must come from the release plan.
+- **`AS_OF` was a hardcoded literal.** `build_catalog_view.py` re-stamped
+  `"2026-06-06"` on every rebuild forever. A generated artifact that lies about
+  its own freshness is worse than a stale one, because it defeats the check.
+- **`/me`'s `track_count` counts PUBLIC tracks only** — 70 reported against 132
+  held. Any tool sizing a walk from it under-counts by 47%.
+- **A generator compared by track number across a re-cut.** The Aug-12 set had 15
+  tracks and v4 has 14, so dropping *Desire* renumbered *REVOLUTION* 15→14 and the
+  manifest confidently reported "#14 Desire, 388.1s vs 205.6s" — REVOLUTION's
+  length. Key on identity, never on position, across generations.
+- **`git push origin master` from a non-master HEAD is a silent no-op.** A
+  concurrent session moved this shared checkout to its own branch between two of
+  my commits; four landed there, and the push pushed the already-current local
+  `master` ref, printed nothing, exited 0. Recovered by worktree cherry-pick. The
+  worktree rule for shared repos has no orchestrator exemption.
+- **Assembling a check revalidates it.** check-mix went from *unrunnable* straight
+  to *failing*: Tidal 07/08/10/12 at -inf dB in the saved session, the third
+  recurrence of that bug. A closed fader and a silent instrument are identical on
+  a post-fader meter.
+
+**Deps.** Closes the diagnosis half of A3 and supersedes its "two known defects"
+note. Blocks on PLN for `--go` and for `/desire`. Unblocks the 8 remaining uploads,
+then OPAL's empty `audio:` field on the website and the catalog release links
+behind it. `sc_generation_lens.py` is reusable and CosmicFest will need it.