Rejecting silent loops at export was the right fix in the wrong place. BIGHEN's
strings stem plays in two sections and rests between them, and all six of the
windows the finder ranked highest sat inside a rest — so filtering at the gate took
`fred_bighen_tonal` from four silent loops down to one real one, when it should
have gone to four real ones. `fred_bighen_fx` disappeared entirely for the same
reason.

The composite is degenerate on silence in the finder too, and for the same reason
it was in the grader: a rest is maximally self-similar in the SSM, its seam is
immaculate, its zero crossings are trivial. So presence is checked before ranking,
on the window itself, sharing `grade.THRESH["empty_dbfs"]` by import so the two
stages cannot disagree about what silence is.

The test for this took two attempts, and the first one was the interesting one. A
stem gated on its own beat grid CANNOT reproduce the bug: `beat_grid` finds no
beats in the silence, so no window is ever generated there, and the test passed
with the gate disabled — proving nothing. The bug needs a grid derived from ANOTHER
stem, which is exactly how `stempack` calls `analyze_stem` and exactly why it does
so: sharing a rhythmic track's grid gives every stem windows across the full
duration, including the parts where that stem is not playing. Verified the test now
fails with the gate off and passes with it on.

Suite 119 → 122.