If you have a subspace of codimension k, that means your convex polytope is defined by some number of inequalities and k independent equalities. So you can still use modified rejection sampling:
- Find n-k independent variables p_1 through p_{n-k}.
- Compute the possible ranges for those variables
- Sample each variable.
- Compute p_{n-k+1} through p_n
- Accept if it's within your simplex, else reject and repeat.
I'm fairly sure this is still uniform because the dependent variables are related linearly to the independent ones and so mumble Jacobian mumble linearity, but I can't quite prove it.