PLN found this by ear, twice in ten minutes, while recording OPAL takes:

  "techno drum mask is inverted on d8, should ALWAYS be the mask midiOff on ^60
   and the ply or other multiplier/nassim button effect on the push-release 92"
  "drums in revolution seem way too fast ... yea def an inversion"

## The bug I shipped

Phase 2 of the surface remap (5910aacc) column-aligned every button reference, and
assigned each orbit's two cells in FIRST-APPEARANCE ORDER. That is role-blind, and
tracks conventionally write the gesture line above the gate line:

    $ midiOn  "^92" (ply "1 <2!3 4>")      -- momentary flourish
    $ midiOff "^60" (mask "t(4,8,1)")      -- latched gate

so the first control encountered — the gesture — took the LATCHING cell, and the
gate took the momentary one. Measured on the diff itself rather than on corpus
history: of the button lines that commit moved, **32 were inverted**, 11 happened
to be repaired, 10 were already wrong, 2 were right both times. 13 files, nearly
all of them in the OPAL setlist.

"Way too fast" is the audible signature: a `ply` on a latching button stays
multiplied after one press. The mirror image is a `mask` that only gates while a
finger holds it down. Neither errors. Neither is silent. Only the ear catches it —
which is precisely the class of bug that needs a machine check.

## The convention, now written down once

    row E (41-44, 57-60), latching   -> GATES    (mask, struct)
    row F (73-76, 89-92), momentary  -> GESTURES (ply, fast, stut, chop, slice...)

A real limit of the grid, found while encoding it: **d1-d3 own only ONE button.**
Their row-F cells (73,74,75) are the per-family mutes gMute1/2/3, so a gesture on
^41/^42/^43 is not an inversion — it is the only cell that exists. Every tool here
skips those columns rather than inventing a slot.

## tools/button_roles.py — one judgement, three consumers

migrate-columns.py ASSIGNS cells, fix-button-roles.py REPAIRS them, pvlint PV010
REPORTS them. Three copies of a musical rule is three chances to drift (the #97
lesson, applied before it could bite). Two parsing rules were earned in the space
of one afternoon, and both were wrong in my first cut:

1. **A gate beats a gesture at the SAME level.** `mask "t(8,16,1)" . chop 16` is a
   latched break-gate that happens to chop while open. Treating it as ambiguous is
   what hid revolution's d8 pair from the first scan — the very pair PLN heard.

2. **But only at the same level.** `superimpose (struct "t . t(3,8)" . arpeggiate
   . (|+ note 12))` is a HELD FLOURISH; the `struct` builds the added layer's
   rhythm two levels down and says nothing about the button. A plain substring
   search called phunk's d6 a gate and moved it to the latching row — the same
   inversion, one level deeper. So the role belongs to the top-level chain, with
   nested arguments stripped and string literals skipped (`mask "t(4,8,1)"` would
   otherwise look like it opened two parens and swallowed the rest).

3. **The body is not the rest of the line.** `midiOn "^91" ( -- SLICE!` puts the
   function on the FOLLOWING lines. A same-line regex returns "unknown", and an
   unknown control gets a positional fallback — re-creating the inversion. The
   classifier now follows the parentheses, capped at 8 lines.

## Validation

- 214 tests pass (24 new: 14 for the classifier, 10 for PV010/PV011).
- The classifier is CONVERGENT: applying the repair twice yields 0 further
  rewrites. That is what caught mistakes 2 and 3 — each fix made the tool disagree
  with its own previous pass, and the disagreement was the bug report.
- pvlint on the setlist: 0 errors, and PV010's 12 unfixable cases are reported as
  `info`, not warnings, because an orbit with two gestures genuinely cannot put
  both on the momentary row. A lint that nags about the impossible gets ignored,
  and this one has to stay trustworthy enough to gate a gig.

## PV011 — one FX bus slot shared by two orbits

PLN's ask, right after PV004 surprised him with a shared cut group: "putain a
unexpected cut shared, well done! like shared busses they screw things. can we
lint and flag these? diff bus shared button is ok, but to be detected too". A
`<fx>bus N` is a SLOT in SuperDirt, so two orbits naming the same number share one
instance and the last event to land sets the amount for both. It immediately found
take_5_drops putting d4's bass and d7's choir on `crushbus 41` — the standing
suspect for "d7 feels way more attacky than before". Different effects on the same
number are fine and are not flagged: the namespace is per effect, exactly as he
said.