A trip points at a shape with shape_id, so one shape can serve every trip
that runs the same way. Some feeds generate a shape per trip, or per tiny
variation of a pattern, and end up shipping the same line hundreds of times
with a few centimetres of noise between the copies. Every consumer downloads
all of them, stores all of them, and draws the same line on top of itself.
The law
At least 2,000 shapes, and at least 15% of all shapes, are copies of another shape: they start and end within the same ~100 m, pass the same point halfway along to within ~1 km, and run the same length to within 5%. shapes.txt is judged on its first 128 MB and the count scaled up to the whole file.
How to get pardoned
Merge shapes that run the same way and point every trip of the pattern at the
one that is left. If the copies differ only in where a trip starts or ends
along the route, shape_dist_traveled in stop_times.txt lets them share a
shape anyway.