Its very hard to follow as I don't know what the labels are for each effect (and I'm not familiar with process*). However, in general:

1) In a mediation model there's no reason why a a direct effect might be reduced to zero if there is a mediation effect. (In fact that's usually the reason people run mediation models)

2) As there are multiple covariates this can change the effects if there is collinearity between predictors (because the effects in the regression model are the unique effects after accounting for variance shared with other, collinear predictors).

3) you may have an interaction effect, if so its possible the effect you are referring to is being evaluated some particular, nonsensical value of the moderating variable (such as 0)

So there are multiple reasons why this sort of thing can occur.

* I don't use SPSS anymore, but in any case I think I'd run these sort of models using a more old-fashioned two-stage approach to check if I used SPSS.

In addition to endorsing the points made by Thom Baguley I would mention that you say your model includes "cognitive processing load which was used in my pre-test to measure the difference in visual clutter." That seems to imply you have two different measures of visual clutter in the model -- not only the 2-category variable you used in the t-test, but also this 7-point Likert item. If that's what you've done, then you have built in a severe collinearity problem (even without the interaction term, which is the product of those same two variables). It does not make sense to evaluate the effect of (2-category) visual clutter while controlling for (7-point) visual clutter. It is therefore not surprising that the effects of interest in that model are non-significant.

P.S. Please correct my parenthetical about the interaction term being the product of the two clutter variables; your acronyms confused me. I think the interaction term involves only the 2-category clutter variable and escapism, not both clutter variables. Nevertheless, the main point of my comment still applies, if I am correctly understanding that (7-point) cognitive processing and (2-category) visual clutter are alternative measures of the same construct.

Master data

Bootstrapping

Thank you for all your answers. To clarify what the output means GHVCLVC means the 2 category independent variable for visual clutter (Grouphighvisual clutter/low visual clutter).

Escapism is my moderator (Mescapismbs) measured on a 7 point likert scale.

Next to this Burke D. Grandjean, I include Cognitive processing load (MCPLBS a.k.a. mean cognitive processing load) as a covariate in the model and yes it is the measure used in the pre-test to test the difference in perceived visual clutter. My supervisor told me that I can include this variable as a manipulation check variable, I understand the issue of multicollinearity but why would one ever add a manipulation check variable into the multiple regression?

Perhaps you misunderstood your supervisor's suggestion. As a manipulation (or validity) check, you could examine the bivariate association between the two variables (which should be strongly associated). But I can't think of a compelling reason to include them both in your interaction model, because of their collinearity. Using only the 2-category measure of clutter should give you more meaningful results.

I suppose you could then add the 7-point measure to the model on a separate step, to show that the results from the previous step become non-significant with that one additional variable controlled. You might interpret the non-significance as a manipulation check of sorts, in that it demonstrates the effects of 2-category clutter disappear when the 7-point measure is controlled; hence all of the effect of the 2-category measure is shared with the 7-point measure. But that seems unnecessarily complicated to me.

Thanks once again Burke Grandjean, I have removed the manipulation check variable from the analysis and this still came up with results that did not make sense to me. Then due to reading this post https://www.researchgate.net/post/Different-results-in-PROCESS-and-SPSS, I just noticed that I did not click the option to mean-centre for bot the dichotomous and continuous variable. This was another crux of the issue and now my results are finally meaningful (see SPSSoutput3 for the results).

On the final note, now I have both a process hayes multiple regression (moderation) output with a dichotomous independent variable, a continuous moderating variable and continuous dependent variable, and I have an ANCOVA output with a dichotomous independent variable and a dichotomous moderating variable ( median split). Which approach would be more insightful for my results section?

If I'm reading your output correctly, the interaction term is not significant (or even close to significant). Assuming the results are similar with either a dichotomous or a continuous moderator, it doesn't matter much which way you choose to report it.

Either way, it seems to me you just report the interaction model briefly, in passing, as an interim result on your way to a simpler model that omits the interaction altogether. With that simpler model, you would be in a better position to assess the effect of MESCP, the non-significance of which in the interaction model could well be due to its collinearity with the interaction term. The principle of parsimony would suggest basing your interpretations on the simplest adequate model, which looks to be the no-interaction model.

If there were an interaction that's significant (or close enough to significant that you can't be comfortable assuming it is non-existent), and if the results are otherwise similar using either a dichotomous or a continuous moderator, it is generally easier to understand and explain the nature of an interaction based on a dichotomous moderator. Then you can say something like, "the effect of X on Y is (b1) when Z=0, whereas it is (b1 + b2) when Z=1" instead of saying, "the effect of X on Y changes by (b2) for each one-point change in Z". The latter is more complex, and harder to visualize or graph.

(BTW, for similar reasons I personally avoid mean-centering a dichotomy, even if other predictors in the model are mean-centered.)