Delayed choices undermines past to future causality only if one assumes that the initial state of a system must directly determine measurement results, and that wave functions represent our ignorance of an underlying state which is already determined and only revealed by measurements, but there are other interpretations.
In sum, the situation is as follows: after a measurement, in retrospective, we can come up with a nice classical story of what happened, but the story would have been different (including for the initial state) had we decided to measure something else. But before the measurement, we cannot know the relevant aspects of the initial state that would tell us which kind of "classical story" will occur. All "classical stories" are still possible, so it's not as if the measurement changes something we knew for sure about the system, that suddenly became false. Measurement only affects something we would know if we had measured the system differently, so it's a problem only if one assumes that this "something we would know" must already be determined in some way from the start of the experiment or in other words, if we insist in having classical stories of experiments.
Here are some more details.
Quantum mechanics puts some constraints on causality but things are a bit complicated.
All depends on the notion of causality.
Some authors, e.g. Lewis, attempt to reduce causal talk to counterfactuals, correlations and the like (A cause B iff A had not occurred, B wouldn't have occurred, or if A rises the probability of B...). Reichenbach's principle can be invoked to that effect (if two events are correlated, either one causes the other, or the two have a common cause).
Causality would be reducible to correlation talk + temporal relations.
Having causal relations of this sort in an experiment is what I called above having a "classical story" of the experiment (the technical term is counterfactual definiteness).
With such an account, you'd have to admit that there are non-local causal influences in quantum mechanics because we find non-local correlations. Since which correlations are displayed depends on the way a system is measured, either the way a system is measured has retrocausal influence on the initial state, or it has instantaneous non-local influence. In the latter case, since according to relativity there is no absolute ordering in time of distant events, it seems that one cannot say anymore whether the cause precedes its effect or not (unless one introduces a privileged frame of reference, which is at odd with relativity). So in any case under this conception of causality we have a problem with time ordering of causes and effects.
However there are other accounts of causality: causality as production (A causes B iff it is possible to produce A to make B occur) or energy or information transfer. Under these accounts there is never a non local influence because there is no way an agent could influence a remote measurement as s/he wishes only by deciding to measure the system in a way or another at his position. This is because in any case, the measurement result will be random. Then the correlation at a distance could be explained not by a causal influence from one place to another, or by a common cause, but for example because the properties we measure pertain to a non-local, holistic system.
So although some aspects will remain puzzling, there are ways to understand causality in such a way that no retrocausal influence is entailed by quantum mechanics.
Your question is about the eraser experiment specifically.
In this case you have a double slit experiment, where the particles are entangled with others which are used at a later time so as to extract an interference pattern from the first experiment. This extraction involves "erasing" the information on which slit the particles went through.
Note however that the pattern initially measured is not changed by future actions: it shows no interference, we only extract an interference pattern from it by selecting some particles (the ones that go through the eraser). Which particles will go through the eraser is unpredictable. So although it seems natural to view the eraser as causing the interference pattern retrospectively, one could as well say that the interference pattern was already there and influenced which particles would go through the eraser or not.
What this means is that adopting a view of causality as correlation, the eraser experiment can be explained by non local causal influences without recourse to retrocausal influences because the pattern is actually extracted (not created) in the future.
The quantum eraser experiment is particularly challenging only if you have a classical picture in mind (it exacerbates how quantum mechanics violates our intuitions) but in effect, it's just a normal quantum experiment where a wave function gets measured in various ways and correlations are observed. In any case no past measurement result ever change and there can only be a problem if one assumes that there must be something out there that determines these results according to a "classical" story.