docs(tasks): archive #93/#106/#107/#31 to the structured completed log

Four tasks closed 2026-07-29 evening: the silent-eval harness fix (and the real bug
it uncovered), the take-pack sidecar+archive tool, and the desire rhythm study.
Written for a cold reader per the archive's own convention — commit hashes are in
git log, this is the narrative and the numbers that don't show up in a diff.

armada/tasks/completed-archive.md

@@ -1037,3 +1037,119 @@ it). **59 knob substitutions across 13 tracks, 5 gSel FIXMEs, 0 collisions.** Co
   ignore. Hence `--knobs`.
 
 **Deps.** #46; blocks #94 (buttons, 166 refs, now unblocked by gMask's retirement).
+
+---
+
+## #93 — silent-eval harness cannot BUILD 3 of 13 setlist tracks
+
+**Description.** Found while verifying the #92 column migration, 2026-07-29. Two of
+the harness's own build failures (ambiguous `cutoff`, ambiguous chord-literal
+`IsString`) had been sitting there, mislabeled as "harness limit, ignore this" — a
+message that turned out to be actively dangerous.
+
+**Done.** Two real harness bugs fixed by matching ghci's actual defaults instead of
+diverging from them: hide every top-level name the generated module defines from the
+`Context` import (in ghci a `let` shadows an import silently; at module top level
+there is no shadowing, so the fix is to make hiding explicit and derive the hide-list
+from the generated text so it can't drift); turn on `ExtendedDefaultRules` +
+`NoMonomorphismRestriction`, both on by default in ghci. 12/13 tracks now build.
+
+The verdict computation changed too: build errors whose GHC line number falls inside
+the track's own generated lines now get a distinct `BROKEN` verdict — "THE TRACK DOES
+NOT COMPILE — it will fail live too" — and count as a gig blocker; errors in the
+harness scaffolding stay `BUILD FAILED — harness limit`. Line numbers are mapped back
+to the real `.tidal` file by matching the offending line's text.
+
+**Learnings.**
+- **The distinction between "our fault" and "your fault" must be computed, never
+  asserted as a constant string.** The harness's own reassuring message was right
+  twice and catastrophically wrong the third time — it told the reader to ignore a
+  real, gig-blocking bug in `desire.tidal` (#108) using the exact same words it used
+  for its own limitations.
+- **Two apparently-separate build failures had the same root cause**: the harness
+  environment diverging from how ghci actually evaluates the same code. Fixing the
+  divergence generically (match ghci's defaults) resolved both, rather than patching
+  each symptom (annotate this one literal, rename that one binding).
+- **The 13th failure wasn't a harness bug at all.** Once the two real harness bugs
+  were fixed, the last one turned out to be a genuine typo committed in
+  `desire.tidal` — see #108.
+
+**Deps.** Found via #92. Unblocks accurate pre-gig verification of the whole setlist
+for the first time. Directly produced #108 (gig blocker found).
+
+---
+
+## #106 — Per-take sidecar: the MIDI performance must travel with the audio
+## #107 — Take archive: wavpack compression
+
+**Description.** PLN, 2026-07-29, recording OPAL practice: *"ensure recordings have
+the midi too, keep this from blowing up tracking either ardour rec or compressing to
+avoid storing hours of no-move?"*. Two asks, one command.
+
+**Done.** `tools/take-lens.py pack <take> [--compress]`. Sidecars a take's MIDI
+(every cc/note/track/mark record in the window, copied out of the gig log before its
+prune timer deletes them), meta (window, per-orbit peaks, xrun delta, gear/boot
+state), and EDL, next to the audio rather than in a directory nothing else protects.
+`--compress` then `wavpack -h`s every source and verifies EACH with `wvunpack -vm`
+before reporting bytes saved. Measured on take 94 (21 min, 12 orbits): **5.84 GB →
+1.25 GB, all verified lossless.** Never deletes the `.wav` — that's a separate,
+explicit, human step after the take has been heard.
+
+**Learnings.**
+- **`xrun` in gig-log's `s` records is a running total, not a per-tick delta** — the
+  same `CUMULATIVE` trap the log format documents in its own code, walked into again
+  by a new consumer. Summing it reported 58,668,469 xruns for a 21-minute take.
+  Rebased (last minus first): 230, a real and useful number.
+- **Records != events.** A `cc` record is a one-second coalesced bucket carrying a
+  count `n`; reporting the record count instead of `sum(n)` understated the
+  performance ~25x.
+- **FLAC was rejected outright, not benchmarked against.** It can't store 32-bit
+  float, and routing through s32 would hard-clip the 8 orbits sitting over 0 dBFS
+  while looking like a clean win on the one silent orbit. A format that can silently
+  destroy the loudest material is not a candidate, regardless of ratio.
+- **Found in passing: the LED daemon was silently mute in its own journal.** Two
+  independent causes — Python block-buffers stdout to a pipe (always true under
+  systemd, fixed with `PYTHONUNBUFFERED=1`), and its own `-q` flag was suppressing
+  the one line that mattered (track changes) rather than the per-CC chatter it was
+  meant to quiet. `quiet` needs to mean "no spam," not "no signal."
+
+**Deps.** Pairs together by design (one `pack` run does both). #106/#107 depend on
+nothing; the trim-threshold fix ([[feedback_trim_threshold_ear_correction]] in
+memory) landed the same evening on the same tool.
+
+---
+
+## #31 — desire.tidal bass rhythmics study
+
+**Description.** PLN, 2026-07-29: *"i gotta solve desire bassline is not great
+rythmics coulld be cooler more nto more ete_a_mauerpark or insouciance or
+haunted_house i guess. letting it cool."* — pointing at his own prior tracks as the
+target feel, then deliberately stepping back from the decision.
+
+**Done.** `study/desire_bass_rhythm.tidal` — read all three referenced `.tidal`
+files, named the current line's mechanism precisely (four equal `.`-groups, one
+repeated note, density-only variation via `*<n>`, no rests, strictly on-grid — "a
+metronome with a stutter"), extracted the six mechanisms the references share and
+desire has none of (uneven durations via `@`, pitch=rhythm, eighth-note calage
+displacement, accent following the rhythm's own weighting, rests, a phrase longer
+than one bar). Wrote four alternatives, each isolating a different subset of
+mechanisms so the reaction is attributable, verified via `silent-eval --seeded` to
+compile and measurably differ (3.0/4.0/4.5/5.75 onsets per bar vs ~5.5 today). Pitch
+design (D minor, measured corr=0.903) held constant throughout.
+
+**Learnings.**
+- **The densest option (D, 5.75/bar) is the control that matters most.** It's
+  *busier* than the current line and still reads as a riff — proving the original
+  complaint was never about note count, it was about where notes fall and how long
+  they last. Without that control, "add some space" would have felt like a guess
+  dressed as a diagnosis.
+- **Different shape from the copycat/algoraoke workflow.** That one matches an
+  EXTERNAL track via web research and demucs. Here the references are the artist's
+  own prior work already in the repo — the method is read the actual source, name
+  the current mechanism precisely, extract the shared grammar, hand back isolated
+  options. See `feedback_self_reference_rhythm_study` in memory.
+- **"Letting it cool" is an instruction, not a delay.** It means prepare options
+  without pushing a decision — the study never touches his dirty working copy.
+
+**Deps.** Depends on nothing. Surfaced #108 as an unrelated finding (desire.tidal
+doesn't currently compile at all — a one-character typo, not a rhythm issue).