Does the item need to be reverse scored? Or is it an item that few participants endorse or can relate to. Essentially, you could examine why that item is functioning differently (if it doesn't need to be reverse scored). Consider item response modeling/DIF testing. 

Dear Robert,

it depends on the Mode used. Usually, composites are build by Mode B and common factors are usually estimated by PLSc using a correction on the estimated weights obtained by Mode A. To my opinion, the loadings (correlation between the construct and the indicators) can have a different sign compared to weights if they are obtained by Mode B. In Mode B, the weights represent partial correlation (regression coefficients). For Mode A weights and loadings should have the same sign, as weights are equal to the covariance between the construct and the indicator, For composites (I try to avoid the term formative, as it has several meanings), assessing discriminant validity doesn't make sense, as the indicators are not measures of the underlying construct (see for example Henseler et al. (2015)).

Best regards,

Florian

Henseler, J.; Hubona, G. & Ray, P. A.

Using PLS path modeling in new technology research: Updated guidelines

Industrial Management & Data Systems, Emerald Group Publishing Limited, 2016, 116, 2-20 

Dear Robert,

I guess you refer to footnote 5. On a first sight, from a theoretical point of view, I see no problem, if you have negative weights and negative loadings for some indicators. I would say if this fits to your theory/interpretation/expectations then it is okay. Changing the coding will change the sign, but I'm not sure if this helps in your case. Sorry for my little helpful comment, perhaps other can help. 

Best regards from Würzburg ;),

Florian