The expression `src_offset_x - coords->x + origin->x` is parsed as
`(src_offset_x - coords->x) + origin->x`; since floating point
arithmetic is not generally associative, even when
`coords->x == origin->x` (in particular, when there is no active
symmetry), it may still yield a different result than plain
`src_offset_x` if there's not enough precision for the intermediary
result (which is usually the case when `{origin,coords}->x` is
noninteger.)  Since `src_offset_x` is an integer, and since the result
of this expression is rounded to an integer, if the error happens to
be in the direction of the rounding, it's magnified to a whole pixel,
which causes visible "jitter".  (Ditto for `src_offset_y` and co.)

Regardless of this issue, we want to individually round `origin->[xy]`
and `coord->[xy]` down before taking their difference, since the
original offset is calculated according to rounded-down coordinates.
This solves the original issue along the way.