docs(tasks): rebuild the board as TASKS_DUMP.md — the Task API is gone

The Claude-Code Task API is not reachable from this session (TaskList/TaskCreate/
TaskUpdate/TaskGet all absent), so the live board — last seen at 112 tasks —
could not be read or rewritten. Rebuilt the tree from durable sources instead:
board-archive.md (ids exact, completed through 2026-07-29), the numbered
achievement logs, the memory store, and git.

Structured as six epics with the work nested underneath: OPAL-26 release (the
only one blocking a shipment), cut/boundary detection, the floor problem, rig &
hardware, open taste calls, completed. Each item carries the measurement or the
quote that justifies it rather than a bare title, so it survives a cold read.

Honest about provenance: ids in the #150+ range come from the lost board via
session context and cannot be re-verified, so the file says to treat a #17x
number as a name, not a key.

This session's outcomes are slotted in place: #1's leading silence measured
(1.452s), the timbral cut lens declared a dead end with its three-attempt table,
and #13/#15 escalated to ear-only calls.

TASKS_DUMP.md

+# TASKS_DUMP — ParVagues / L'Armada board reconstruction
+
+**Written 2026-08-16.** The Claude-Code Task API is not reachable from this session
+(`TaskList`/`TaskCreate`/`TaskUpdate`/`TaskGet` all absent), so the live board — last seen at
+112 tasks — could not be read or rewritten programmatically. This file rebuilds the tree from
+durable sources instead, and is meant to be the thing you edit or re-import from.
+
+**Provenance of every id below**
+
+| source | trust |
+|---|---|
+| `armada/tasks/board-archive.md` | verbatim, ids exact (completed through 2026-07-29) |
+| `armada/tasks/NNN-*.md` (001–025) + dated logs | verbatim, narrative achievement logs |
+| memory store + this session's context | ids **recalled**, not verified — treat as labels |
+| git log | authoritative for what actually shipped |
+
+Ids in the #150+ range come from the lost board via session context. They are almost
+certainly right but I cannot re-read them, so **treat a #17x number as a name, not a key**.
+
+---
+
+## EPIC A — OPAL-26 release  ⟵ *the only epic blocking a shipment*
+
+North star: get "Sunset Forest" (opal-festival-2026, 2026-08-08) released. 14 tracks —
+**#14 Desire is cut**, PLN confirmed 2026-08-16; #13 Vague de CRIME becomes the last official
+track, #15 REVOLUTION stays the Encore.
+
+### A1 · #177 — Re-cut v4 from ear-verified boundaries  🔴 IN PROGRESS · blocks everything below
+
+Ground truth: `armada/tide-table/judge_specs/opal26_boundaries_ear.json` — **do not overwrite
+from a detector.** Nine boundaries verified by ear (#3 893.0 · #4 1118.4 · #5 1549.4 ·
+#6 1808.1 · #7 2248.8 · #8 2726.4 (alt 2730.9) · #9 2954.4 already correct · #10 3269.0).
+
+- [x] **#1 leading-silence measured** — 1.452 s of *literal digital silence* (all samples zero
+      to sample 64038; first transient −28 dBFS). Objective, no ear needed.
+- [ ] **A1.1 Apply the #1 head trim** — ⭐ fully independent of everything else, do it first.
+- [ ] **A1.2 Settle #11, #12** — candidates exist (`cut_candidates_v3run2.json`), need one ear call each.
+- [ ] **A1.3 Settle #13, #15** — 🚨 **ear-only, no candidates will ever come from the timbral lens** (see B1).
+- [ ] **A1.4** Write `segments_v4.json` → `v4_provenance/finish.sh` → rebuild judge data.
+
+### A2 · #183 — REVOLUTION mix fix  · blocked by A1
+
+- [ ] Cut **d5 `the_revolution`**, *not* d10. PLN's first guess was wrong: measured
+      band-limited (400 Hz–8 k, what a laptop reproduces) d5 leads d10 by **13.7 dB**;
+      cutting d10 changes nothing he can hear. See memory `reference_small_speaker_and_rights`.
+- [ ] Root-cause **d9 SUGAR CHOIR = digital silence** (−240 dB) while declared in the score.
+      Two independent measurements agree. This is the finding that motivated D2's "shout until played".
+- [ ] Measure the bass-"saturation" claim separately — it is a different question from loudness.
+
+### A3 · #179 — SoundCloud uploader (Playwright)  · blocked by A1 · **NOT BUILT**
+
+Designed and specced only. PLN's ruling: *"bite the bullet and use full browser automation…
+we cant secondguess api, nor limit ourselves to half auto process."*
+
+- [ ] Build it. Only `/uploads/*` is DataDome-walled; **metadata PUT is free** — see
+      `reference_soundcloud_write_path`, `reference_soundcloud_cli`.
+- [ ] **Uploaded ≠ cleared.** Content-ID flags arrive by *email*, never through the API, so the
+      tool must model those as separate states.
+- [ ] Destructive calls must pass an allowlist predicate on a test-namespace permalink prefix.
+      The 110 real tracks must be unreachable **by construction**.
+- [ ] Delete playlist `2282681433` — sanctioned by PLN, **only once v4 is up**.
+
+### A4 · Rights / Content ID
+
+- [ ] **#4 "Am i Doing it Right" is flagged on SC** and it is a *true* positive — the score
+      literally uses the `daft` sample bank. The durable asymmetry: the same audio **passed
+      inside the 87-minute continuous mix** and flagged as a 7-minute standalone.
+- [ ] Consequence to act on: static-check every track's `.tidal` sample banks **before**
+      planning a per-track release. The full-set upload is the lower-risk path.
+
+---
+
+## EPIC B — Cut / boundary detection
+
+### B1 · #178 — Crossfade-blind cut detection  ⚰️ **DEAD END, declared 2026-08-16**
+
+Three attempts, all validated against the ear and all recorded in-code with their numbers:
+
+| | approach | median err | coverage |
+|---|---|---|---|
+| `cut_lens.py` | midpoint reference, weighted mean, threshold | 24.0 s | — |
+| `cut_lens2.py` | adjacent refs, median over orbits, CUSUM | 16.6 s | 8/14 |
+| `cut_lens3.py` run 1 | span + self-calibrating fraction `u` | 3.7 s *(n=2)* | 8/14 |
+| `cut_lens3.py` run 2 | medoid refs + effect-size gate | 0.4 s *(n=2)* | **4/14** ↓ |
+
+**Do not tune this a fourth time.** The accuracy figures are not evidence — n=2, and those two
+boundaries (#8, #10) already had ear answers; a metric computed on the cases that were already
+easy measures nothing. The real result is the pattern: **#13 and #15 produced zero voting
+orbits across all three gates and two estimators.** Those transitions have no orbit whose
+timbre changes materially across the switch — the outgoing and incoming tracks share their
+sounds — so a timbral lens is structurally blind there, exactly as an activity lens is blind
+to a crossfade. Loosening further can only admit noise.
+
+Two ways out, if ever revisited:
+- [ ] **Change the feature** — harmonic/key change, or per-sample identity from the score's
+      sample banks (which differ even when the spectrum does not).
+- [ ] **Accept ear-only** for the blind cases. Cheap and honest: two timestamps, seconds each.
+
+Standing rule: validate any new attempt against `Montreuil26_master/boundaries_ground_truth.json`
+(14 user-confirmed) **first**. Note Montreuil has no per-orbit stems, so a stem-based lens
+cannot use it — that gap is itself worth fixing before attempt four.
+
+### B2 · Boundary webview  · unblocked, not started
+
+- [ ] Per-cut transition view: ±45 s waveform, markers at nominal/onset/takeover, click-to-
+      audition, export `boundaries.json`. Reuses `WaveformPlayer` + precomputed peaks.
+      With #13/#15 having no candidates this is a **plain audition-and-mark tool**, not an A/B —
+      simpler than originally scoped.
+
+### B3 · The set-judge  ✅ SHIPPED
+
+`9c6a24c`, `20b9151`, `73e9d79` · log `armada/tasks/025-fifteen-intros-are-not-a-set.md` ·
+memory `reference_set_judge`. Reusable per-gig via `judge_specs/<gig>.json`.
+
+---
+
+## EPIC C — The floor problem  ⟵ *the highest-value non-release thread*
+
+Origin, PLN verbatim: *"I DIDNT KEEP THE KICK LONG ENOUGH. MAIN criticism / negfb I GOT:
+'destructure. les gens etaient un peu perdu'."*
+
+**Measured, and the data agrees with the floor: the kick was audible 59% of OPAL-26.**
+13 of 15 tracks under 70%. Longest kickless stretch 164 s (#4). `f1798e7`.
+
+### C1 · #180 — Pre-gig floor lens
+- [ ] Static + stem analysis that flags a track's kickless stretches **before** the gig.
+- [ ] **Count BARS, not seconds** — the set ran 89–166 BPM, so a seconds budget is wrong by 2×.
+
+### C2 · #181 — HUD v2: debts + instruments  🎯 *the design is settled, see memory*
+PLN: *"HUD needs evolve from 'just a lcxl repr' to proper parvagues success metrics/tooling/
+state helper"* · *"things should shout at me until playd"* · *"2am-elevated-UX … like
+macdonalds' screens help avoid cold burgers"*.
+
+One mechanism — **DEBTS**, things that accrue until discharged:
+- [ ] **floor debt** — bars since the kick last sounded. Must be a *budget* (≈16 bars calm →
+      amber → red past 32) with an explicit "holding a breakdown" override. **A drop is
+      supposed to remove the kick**; an alarm that fires on every intentional breakdown is
+      ignored within one set, and then it is worse than nothing. This is the make-or-break.
+- [ ] **unplayed debt** — an orbit declared in the buffer that has produced no events since
+      the track loaded. Shouts until it plays. (Born from d9 SUGAR CHOIR, A2.)
+- [ ] Two registers: **alarms** at the screen edge, colour **+ motion**, no numbers; **instruments**
+      (kick % this track / this gig) in the topbar, never blinking.
+- [ ] **Signal source**: the buffer parse structurally cannot answer any of this — it only knows
+      what was *written*. Add an OSC target aimed at the HUD (`BootTidal.hs` already defines
+      57120 / 1818 / 6013, so this is a one-line change avoiding contention). One stream answers
+      all of it. Prior art `tidal_ears/osc_monitor.py`. **Verify the cost first** —
+      `feedback_observer_perturbs`.
+
+### C3 · #176 — `^41` standardisation across the corpus
+- [ ] Verified from source: `midiOn ch pat = someCyclesBy (orDef 0 ch) pat`. So `^41` is a
+      **fraction-of-cycles** control, *not* an on/off mute, and `orDef 0` means an untouched
+      `^41` is safe (no mute-bomb). A mid value like 64 = the kick on ~half the cycles —
+      literally structure that comes and goes, so **`^41`'s value is a first-class HUD signal**.
+
+### C4 · HUD touched/untouched + scenes  · plan file exists, unstarted
+Plan: `~/.claude/plans/zazzy-honking-waterfall.md` (touched-state tracker via `aseqdump`
+read-only MIDI tap + perf-isolated scene backdrop). **The law: the HUD must never be able to
+stall the editor or audio.**
+
+---
+
+## EPIC D — Rig & hardware
+
+- [ ] **#154 — USB hub is a single point of failure** (drops interface + LCXL together).
+- [ ] **LCXL v3 ordered** (2026-08-16). Retires: pot-pickup value jump (endless encoders send
+      relative deltas — *the whole class disappears*), much of the LED-signalling cost (OLED
+      renders values natively), and the USB SPOF (**DIN path is independent of USB**).
+      🔍 **UNVERIFIED and it matters:** can third-party SysEx paint *arbitrary* text to the
+      OLED? Documented behaviour is only "shows parameter values". C2's alarms are **derived**
+      state, not control values — if the OLED is closed, the alarms stay on screen.
+- [ ] CC numbering + relative-encoder mode will differ from v2. Change it in `lcxl_grid.py`
+      and **regenerate** — never in a copy (`feedback_one_parser_per_concept`).
+
+## EPIC E — Open taste calls  ⟵ *waiting on PLN's eyes/ears, not on work*
+- [ ] **#11** LED brightness · **#70** cut groups. See memory `project_open_taste_calls`.
+
+## EPIC F — Completed
+Full tracker history in **`armada/tasks/board-archive.md`** (#1–#103, through 2026-07-29).
+Narrative logs in **`armada/tasks/001-*.md` … `025-*.md`** + dated logs. Structured long-form
+archive in **`armada/tasks/completed-archive.md`**.
+
+---
+
+## Recommended order
+
+1. **A1.1** — the 1.452 s head trim. Independent, measured, done in minutes.
+2. **B2** — the audition-and-mark webview, now simpler than specced.
+3. **A1.2 + A1.3** in one sitting at that webview → **A1.4** → unblocks **A2** and **A3**.
+
+C2 is the highest-value thread once the release ships, and its design is already settled —
+it needs building, not deciding.