What to do in regression when an important continous covariate is correlated with the grouping covariate? Hoping I can explain my problem clearly enough. 
My Masters thesis is looking at the impacts of elephant on woody vegetation in relation to distance from waterholes and vegetation type. Vegetation type is a categorical variable with 2 levels (teak and terminalia). A regression model might thus look like this.
lmdamage<-lm(damage~distance+propteak+vegtype, data=veg)  

Distance refers to distance from the waterhole. However you will see I have a predictor for the proportion of Teak in a sampling plot, a species of tree which is avoided by elephant. The logic is that damage is lower where teak is more dominant in plots categorised in the 'teak' vegetation type which is borne out by the data. Hence I want to account for this in examining the effect of distance from waterholes on damage. 
When I run the model, all 3 predictors are strongly significant.However, my concern is that in the 'Teak' vegetation type, the proportion of teak varies from very low to 100%, while it is usually 0 in the 'Terminalia' veg type and never more than 10% (I know % teak is sometimes 0 in the teak veg type too, but that's not important for the current problem). Hence I am worried that proportion of teak and veg type might be too highly correlated to use together, or at least that a reviewer might raise such a concern? The proportion of teak is important for explaining damage levels in the teak woodland plots, which is most likely to be a function of both distance from water and proportion of teak, so I do wish to retain it if possible. See the plot for a visual representation of the problem and relationships.
Edit: Added a second plot which shows elephant damage in relation to distance to water and vegetation type


Is there a way to run the model such that the proportion of teak only predicts elephant damage in the 'teak' vegetation type and not the 'terminalia'?
Interestingly the     car::vif     values for the model above are all below 2. However I'm not sure whether these would capture collinearity between a categorical and continous variable? However, if this means I can ignore the problem I just described then great.
 A: A few thoughts on your research problem:
From your description, it sounds like damage in these stands might have more of a direct relationship with other vegetation in the stands than teak. With this in mind,  propteak in your model is a variable that you might expect to have a negative coefficient, as with increasing teak you would expect a decrease in elephant visitation and therefore damage. One option might be to consider including an interaction term between proportion of teak and vegetation type (e.g. propteak:vegtype), as you mention that you believe there is an important relationship between the two.
You might also want to consider if you would expect there to be an interaction between distance and propteak. That is, are the elephants more likely to damage a site with a lot of teak if it is immediately adjacent to a watering hole, than say a site with little teak but much farther away?
And related to this - does the elephant really care if a site is classified as 'teak' or 'terminalia'? If you are including both propteak and vegtype in your model and having both come up as significant, I would ask myself: what else is vegtype representing that would be of interest or importance to the elephants? (Based on the figures you've included, it doesn't appear that vegtype is directly coding whether or not teak is present, so it must be incorporating other information.)
Lastly, although not mentioned in the question (and I might be off in suggesting this), but it might also be good to consider the form of your response, or the units associated with it. If damage is constrained to be greater than zero, or if it is recorded as a percent, you might be better suited to consider a glm() rather than a lm().

Just another quick comment on multicollinearity:
If there is multicollinearity, you are not directly violating any regression assumptions - it doesn't really ruin your inference, except with respect to the interpretation of your beta-coefficients. You can still do t-test, etc., and they will behave as they should; but the problem comes in the interpretation of your individual parameter estimates. With multicollinearity you can't really talk about the effect of one variable (e.g. propteak) while holding all the others constant (as vegtype is correlated with propteak). While you can't separate out these individual effects, you can still do joint tests when trying to consider explanation (e.g. $\beta_{propteak} = \beta_{vegtype} = 0$).
