Remove incorrect empty anchor justification

[jklaise]

The second point in the current version of the docs should not be possible because if it was (empty anchor returned and the prediction always changes, regardless of perturbation), the empty anchor would have precision ~0 and hence would not be returned in the first place. In fact, if the prediction does change a lot regardless of the perturbation, the data point is likely near the decision boundary and the algorithm will try longer and longer anchors to find something that satisfies the precision threshold.

[codecov]

Codecov Report

Merging #701 (94c84b2) into master (e5814a8) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #701   +/-   ##
=======================================
  Coverage   80.55%   80.55%           
=======================================
  Files         105      105           
  Lines       11790    11790           
=======================================
  Hits         9497     9497           
  Misses       2293     2293           

[RobertSamoilescu]

Looks good to me!