This bug is due to the way in which javac calculates glb. When javac computes glb(T1, T2 ... Tn), basically the compiler looks for a type Tk, 1<k<n, such that Tk is a subtype of each Ti, 1<i<n.
In the example however, we have that W = glb(C, I, I1, I2), so javac looks for a type which is a subtype of both C, I, I1 and I2. Unfortunately among the upper bounds of W there is not such common subtype (the situation will be different with C implementing both I, I1 and I2). So the inference fails and the compiler reports the problem.
The JLS state that (5.1.10) "... where the glb(V1, V2 ... ,Vm) is the (intersection) type V1 & V2 ... & Vm. So in this case javac should infer the intersection type C&I&I1&I2 for W. As a general rule, such an intersection type should be created whenever is not possible to find a suitable common subtype among the arguments of glb.