In short, it's not really worth obsessing about how _long_ the course
is, nor even whether the start and finish lines are exactly
perpendicular to the course. But: the start and finish lines _must_ be
parallel, to a very very precise amount. And that's not easy: surveying
(the old-fashioned way) the Finish from an arbitrary Start line to 2
parts per million is not easy. The most up-to-date bit I know of is the
Nikon laser surveying gizmo, which has a range of 2700m, claiming
accuracy of 3ppm (or 6mm/2000m) - but what about older courses? Are they
ever re-surveyed?

You may think you could use a bearing. However, imagine a perfectly
east-west course at 55North (somewhere in England) with the Start and
Finish lines both running due North. Take a guess how much shorter lane
1 will be than lane 6 .....

... 3cm. Just because of the shape of the planet.

(This is why no-one ever wins in lane six)

In ordinary layout the start finish lines would be set 90 degrees off the
course. A true North bearing would not be easily available while lines set
at 90 degrees off the course would calculate to North but with each North
line parallel to the other North line at opposite ends of the course. So
that is grid North or project North and not true North.

Also note that in simple layout diagonal distances can be checked...

Of course in modern layout control points are set near required points and
then the control points located by a loop survey where each control point is
occupied in the traverse loop. Then required points are set at short
distances from the control points.

Do this in water ? Well of course build piers which is not unheard of in
surveying since geodetic surveys build towers...

(I believe that surveys across the English Channel used oil tankers. In this
case the control points in the water are just temporary to get permanent
control points on each shore.)