This graph was determined in the case of simply-laced root systems by Igor Dolgachev and Norman Goldstein *here*. For other root systems the original question should be modified, leading to a precise conjecture which remains unproved. It takes some preparation to state.

Briefly, the fibers in the Springer resolution of singularities of a unipotent class in a simple algebraic group (say over $\mathbb{C}$ or a field of good prime characteristic) are projective varieties, arising from the fixed points of a unipotent element on the flag variety $G/B$. (It's equivalent to study nilpotent orbits of the Lie algebra.) The varieties, as well as their irreducible components (all of equal dimension), have known dimensions (Springer, Steinberg, Spaltenstein), but it remains challenging to describe geometrically the fibers or the patterns of component intersections. The best understood nontrivial case involves the subregular class, where the fiber is a "Dynkin curve" having copies of $\mathbb{P}^1$ as components. The incidence graph in types $A,D,E$ is just the corresponding Dynkin diagram, which coincides in the subregular case with the associated "Kazhdan-Lusztig graph". The latter is defined at the end of the 1979 KL *Inventiones* paper and has one vertex for each component, with vertices joined by an edge just when the components intersect in codimension one.

Dolgachev and Goldstein studied the opposite extreme of the minimal (nontrivial) unipotent class and worked out its KL graph for types $A,D,E$. Their theorem is that *the graph is the same as that for the subregular class*. Their study of the minimal class for $G_2$ yields a different graph. Their consultation with Spaltenstein reported in the paper suggests substitution of the minimal *special* class. (I asked Dolgachev about that a year ago when he was here, but he hadn't seen it pursued further.) The special classes came up in Lusztig's work on Springer representations of Weyl groups, but have no geometric characterization so far. Anyway, Lusztig-Spaltenstein duality for the partial ordering of special classes (generalizing transpose duality for partitions in type $A$) leads in types $B, C, F_4$ to a minimal special class defined by the short root groups, but in type $G_2$ to the subregular class. Comparing data for the *Langlands dual* types
$B, C$ suggests strongly this formulation:

CONJECTURE: In all types, the KL graph for the subregular unipotent class agrees with the graph for the Langlands dual minimal special unipotent class.

A bit of numerical evidence in favor: the number of components of the Springer fiber for the minimal special class in Langlands self-dual type $F_4$ is 6, as for the subregular class of $F_4$, whereas for $B_3$ and $C_3$ the relevant numbers 4, 5 get switched and similarly for types $B_n, C_n$ in general via Springer theory. (For other pairs of classes related in this way, more than a graph would be needed to get a good comparison. Another problem.) Aside from wanting more geometric evidence related to the stated conjecture, my question here is:

Is there a good reason for Langlands duality to play any role in this essentially geometric question?

[EDIT: The paper is now freely available online; link added. But I haven't yet found an answer to my question.]

[UPDATE: Concerning the conjecture itself, an MIT graduate student Dongkwan Kim has followed the suggestion of Roman Bezrukavnikov below to provide a proof *here* using the notion of folding. (It then seems plausible to look for similar behavior whenever two *special* nilpotent orbits are related by Lusztig-Spaltenstein duality.) But my question seems to remain open.]