The image diffs look worse? (a tick that I think we want went away)

Specgram doesn’t include 0 in its extents (its 0.2 or something). Should we include ticks that are out of range? If so how far out of range? I’d really argue this is specgrams decision not the decision of the tick locator, but maybe we had the idea that a tick should be made available slightly outside the actual data range? If so how far outside the data range do we want to allow? Is that where all this distance-along-pixel stuff came from?

Oh sorry, I didn't realize that. In which case I definitely agree with not to put the tick (I think the distance-along-pixel was there for this reason, but only correcting floating point errors).
In fact running the specgram example as of master showcased another funny issue: the tick at 0.0 is present when my window is at its default (rcparam default) size, but disappears if I maximize the window...

It sounds quite natural to only return those tick locations that are within the (closed) interval of [vmin,vmax]. However for some reason there are currently more tick locations returned. I reckon there must be a reason for this; possibly this is a workaround for some rendering problem? In any case I guess it makes sense to find out the reason and then make sure the trunkated version proposed here would not produce the original problem again.

The use case I know of where you want more ticks is automatic levels forcontour where you want the levels to bracketvmin andvmax. So I kept a kwarg for this casetrim_outside=False

Classic-mode autoscalingrequires having the ticks outside the range. The autoscaling determines the range such that lower limit is the largest tick location less than or equal to the data vmin, and similarly for the upper an vmax. The original floating-point slop fiddling, later expanded to work in pixel space, was mainly related to this. For example, when the range is large enough that the ticks fall on integers, one doesn't want tiny and accidental floating point deviations from integers to affect the view limits and ticks.

This looks like the origin of the half-pixel adjustment:255f213

Classic-mode autoscaling requires having the ticks outside the range. The autoscaling determines the range such that lower limit is the largest tick location less than or equal to the data vmin, and similarly for the upper an vmax. The original floating-point slop fiddling, later expanded to work in pixel space, was mainly related to this. For example, when the range is large enough that the ticks fall on integers, one doesn't want tiny and accidental floating point deviations from integers to affect the view limits and ticks.

OK, for classic auto scaling we can easily just use the "trim_outside" flag

This PR includes a tolerance based on a fraction of vmax-vmin to account for floating point slop. I set it unscientifically at 1e-10, but we could do something larger to let it have more slop.

The original issue in#1310 was a floating point mismatch with >. While I admit making the tolerance one pixel fixed the floating point error a) it means we have to do some transforms to figure out that size in data space and b) it means the chosen ticks depend on the pixel size. That probably is not practically a problem, but is not very good theoretically.

The approach here still scales by the data range but just chooses a small tolerance that is very unlikely to be visible, but is well within floating point tolerance

Looking at the code inaxis it seems to me the proper thing to do is to maketransforms.interval_contains actually be floating-point robust....

I took out theinterval_expanded stuffon master out ofaxis.py (i/e/ w/o this PR), and it all passedtest_axes except for the specgram examples. So, I kind of think its obviated (maybe a 64-bit thing?) I think

1. that code should go (a bunch fewer transforms!)
2. interval_contains should be made floating point robust.
3. MaxNLocator should only return ticks between vmin and vmax according to the robustinterval_contains

I think this PR is obsolete now...