Problem

Sometimes, it would be useful if we could "fake" a text's metrics to pretend it has a different physical size than its real size. Consider for example the following example; pretend the two subfigures are actually two separate figures that I want to save an embed in a presentation on consecutive slides, so that the orange line shows up on the second one:

from pylab import *
fig = figure(layout="constrained")
sfs = fig.subfigures(2)
ax0 = sfs[0].add_subplot()
ax0.plot([1, 2], label="foo")
sfs[0].legend(loc="outside right upper")
ax1 = sfs[1].add_subplot()
ax1.plot([1, 2], label="foo")
ax1.plot([2, 1], label="a longer label")
sfs[1].legend(loc="outside right upper")
show()

As it is, doing an animation would result in everything moving around in a rather ugly manner.

Because I know a bit about the matplotlib text internals :-) I know how to temporarily patch matplotlib to fix this, the trick being to swap out the text when matplotlib is measuring its extents, but not when drawing it:

diff --git i/lib/matplotlib/text.py w/lib/matplotlib/text.py
index 40cd8c8cd6..a042555ac1 100644
--- i/lib/matplotlib/text.py
+++ w/lib/matplotlib/text.py
@@ -363,6 +363,8 @@ class Text(Artist):
         """
         thisx, thisy = 0.0, 0.0
         lines = self._get_wrapped_text().split("\n")  # Ensures lines is not empty.
+        import sys
+        if lines == ["foo"] and sys._getframe().f_back.f_code.co_name != "draw": lines = ["a longer label"]
 
         ws = []
         hs = []

which results in

but obviously that is not a "practical" solution (even though I'm happy to have that in my toolbox).

Proposed solution

At least for mathtext strings, we could support the TeX macros \smash, \phantom, and their variants, which are the TeX approach for fixing such alignment problems (see e.g. https://www.tug.org/TUGboat/Articles/tb22-4/tb72perlS.pdf). It is not entirely optimal to only support this for mathtext strings, but that's the best I can think of right now -- adding an additional "faked_size" (or similar) property on Text objects seems less convenient, e.g. for the legend case above.