docs(tasks): archive #70/#105/#108/#109 — the invariant that was a hypothesis

armada/tasks/completed-archive.md

@@ -1153,3 +1153,116 @@ design (D minor, measured corr=0.903) held constant throughout.
 
 **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).
+
+## #108 — GIG BLOCKER: desire.tidal does not compile
+
+**Description.** `live/collab/raph/desire.tidal:46` carried `# pan 0.42plz /se` — a
+committed typo, not a working-tree accident. The file has no blank lines, so Pulsar
+treats it as ONE do-block: that single stray word killed every orbit in a Finale-slot
+setlist track, six days out from OPAL.
+
+**Done.** Fixed by PLN in his own buffer; verified 2026-07-30 (`grep plz` → 0 matches)
+and committed as part of ce887b7. `silent-eval --seeded` now reports **13/13 ok** — the
+first time the entire OPAL setlist has built cold.
+
+**Learnings.** The bug was invisible for as long as it was because silent-eval printed
+every build failure as `BUILD FAILED — harness limit, NOT a track verdict`, a *constant
+string*. It was right about that twice (two genuine harness bugs) and catastrophically
+wrong the third time, telling the reader in reassuring words to ignore a broken setlist
+track. Fixed by computing the verdict from which line GHC pointed at rather than
+asserting it. **Any "don't worry, this is our fault not yours" message a tool prints
+must be reached by evidence** — the dangerous side to be wrong on is the side that says
+"ignore this", because that is the side nobody re-checks.
+
+**Deps.** Found by #93's harness repair. Unblocked nothing formally; it was simply the
+difference between a set that plays and one that doesn't.
+
+## #70 — you_my_sunshine: d5 and d11 share cut group 5
+
+**Description.** pvlint PV004 flagged, on its first real run (2026-07-28), that d5 (the
+vocal) and d11 (`no_sunshine:4/4` chopped) both carried `# cut 5`. Cut groups are
+monophonic, so each d11 event killed the vocal mid-word.
+
+**Done.** 2026-07-30, commit 27c1581 — d11 → `# cut 11`, following the
+group-N-equals-orbit-N convention d3/d6/d8/d9 already used in that same file. Left
+reversible in one edit, with the reason in a comment, because "both vocal layers are
+deliberately monophonic" was a legitimate arrangement PLN could have wanted.
+
+**Learnings.** The find worth keeping is not the fix but the *correspondence*: d5's line
+59 carried PLN's own `-- FIXME VOICE OFF`. **Same bug, two labels** — one filed by a
+linter, one written by ear, months apart, neither connected to the other until both were
+read side by side. They did not look like duplicates in a task list because **the ear
+names the SYMPTOM and the linter names the MECHANISM.** Generalisable: when a lint
+finding and a human FIXME live in the same file, check whether they are one defect before
+treating them as two items.
+
+**Deps.** Found by pvlint (#100's rule set). Closed one of #111's Group-C markers as a
+side effect.
+
+## #105 — Align the gM family: mutes by ROLE, filters by BLOC
+
+**Description.** The corpus treated `gM<N>` as a mirror of `gF<N>` — a legacy of gM once
+being `gMask . gMute<N>` before gMask was retired. PLN wanted the kick mutable on its own
+button.
+
+**Done.** 2026-07-30, commits 72943be (tooling) + 27c1581 (corpus). **917 lines across
+213 files**, corpus-wide rather than setlist-only at PLN's correction (*"not only in the
+set am i right?"*), plus 7 setlist orbits that had NO mute given one (*"no mute is
+oversight!"*). Map authored once in `tools/lcxl_grid.py` (`_FILTER_FAMILY` /
+`_MUTE_FAMILY`), migrator `tools/fix-mute-roles.py`, guarded by new pvlint PV012/PV013.
+Non-set missing mutes reported, not written: `docs/2026-07-30-orbits-without-a-mute.md`.
+
+**Learnings.** Three, and the first is the expensive one.
+
+1. **The task's own stated invariant was wrong.** It was framed as "align gM index to gF
+   index". I measured 35 set mismatches against that rule, rigorously, and was ready to
+   migrate. PLN's actual map groups filters and mutes DIFFERENTLY on purpose — you sweep
+   the drums as a bloc but drop the kick alone — so the real edit count is 55, and
+   `d4: gF2+gM3`, which the mirror rule called broken in 11 of 13 tracks, is CORRECT.
+   Shipping it would have moved 11 good lines and left the kick sharing a button with the
+   hats: the exact complaint that reopened the task. **A stated invariant is a hypothesis
+   about intent, and a rigorous measurement against the wrong invariant produces a
+   confident number answering the wrong question.**
+2. **The linter audited the migrator, immediately.** PV013 was written to prevent future
+   drift; its first catch was the migration in progress. `fix-mute-roles.py` had its own
+   orbit regex requiring `^dN $` and had walked straight past `d5  -- The Voice of Love`
+   (no `$`, gates on a commented continuation). It now imports pvlint's parser. The miss
+   had hidden 14 rewrites and 7 missing mutes — **under-application, not corruption,
+   while every check read green**: convergence said 0 remaining, silent-eval said 13/13,
+   222 tests passed. Verifying the output cannot catch a tool that can't see part of its
+   input; only a second independent parser can.
+3. **The verification that actually answered PLN's "ensure all tracks playable".** Not
+   spot-checks: (a) convergence — run twice, second pass proposes 0; (b) a structural
+   proof diffing all 917 changed lines against HEAD and asserting both sides are
+   identical once mute tokens and whitespace are stripped — 908 provably token-only, the
+   other 9 individually enumerated and accounted for; (c) `silent-eval --seeded` over all
+   213 touched files on a pre-migration worktree AND the migrated tree, **verdict lists
+   byte-identical**. That last one is the real answer, and it cost ~10 minutes of
+   unattended compute.
+
+   It also surfaced #112 as a by-product: 14 corpus tracks that do not compile at all,
+   pre-existing, none in the setlist.
+
+**HONEST LIMIT, recorded because a green check here misleads.** Every button CC seeds to
+0 and `gMute<N>` is `midiOn "^7N" (mask "f*16")`, so all three mutes are equally inert
+until pressed. None of the verification above says anything about behaviour UNDER a
+press — the only thing this change alters. Left for PLN's fingers.
+
+**Deps.** Closed #70 in the same diff. Feeds #44 (`--check` as a pre-gig gate). Related
+to #54/#98, the same "the surface should be one sentence" thread.
+
+## #109 — Stale orphan-orbits fixture (you_my_sunshine d7→d6)
+
+**Description.** `tools/tests/test_orphan_orbits.py:201` expected d7; the parser said d6.
+
+**Done.** 2026-07-30. Fixture corrected to `{1,2,3,4,5,6,8,9,11}`. Cause was PLN's
+d7→d6 consolidation (with ^91→^90, ^59→^58), the same move he had made in do_it_right.
+
+**Learnings.** Re-derived the set by reading the file's column-0 declarations and
+checking for commented-out ones, NOT by pasting the parser's output. **A hand
+measurement that quotes the thing it checks is a tautology** — the whole value of the
+fixture is that it is an independent witness. Cheap to get wrong, and the wrongness is
+invisible forever after.
+
+**Deps.** Blocked nothing; it was the difference between a suite at 213/1-fail and a
+clean baseline to migrate against, which is why it went first.