A gesture is grouped by pitch alone, so it can bridge a silence of any length

[eolivelli]

Found reviewing #663, which made the octave fold decide a gesture rather than a
note.

MelodyEstimator.oneGesture groups notes by pitch alone: two notes that follow
one another in the estimate and lie within a whole tone are one gesture,
however long the silence between them. So a note can be kept in its own octave
by a neighbour tens of seconds earlier — the review measured single gestures
bridging silences of well over half a minute on the committed packages, and
built a fixture where a leakage note keeps a later note a whole tone away from
being folded across three quarters of a minute of nothing.

Nothing in either corpus or in the four field recordings is harmed by it today:
the notes that share a gesture across a long silence are ones the fold would
have left alone anyway. It is not only a missed correction either way — a
gesture takes one shift, so a bridged note can also be carried by a
representative it has nothing to do with, and land where the fold would have
refused to put it.

What a silence condition would cost, measured both ways. Requiring the
notes to touch — a silence no longer than the shortest thing that can be a note
— costs pop-axis read as given the whole of what the gesture rule buys it,
because the notes of a phrase in a mix are parted by unvoiced stretches longer
than any note; that is the only benchmark that moves on those rows. Read
through the separated vocal, which is the path the stage takes, the same
condition keeps melody-level4's gain whole and costs pop-axis part of its
own. Half a second costs nothing at all read as given and the same part of
pop-axis through the stem. So the repair is cheaper than a first reading of
the read-as-given rows suggests, and it is not free.

What would settle it is a rule that reads the signal between the two notes
rather than a length.