Generic type inference for extension methods in C#

I've been coding some exception- and type-safe retrieval-extensions for various generic collections when I stumbled acrross - at least for my understanding - odd behaviour of the c# compiler.

Let's say we have:

var sgDict = new Dictionary<string, object> { ["a"] = "x", ["b"] = 6 };

These two extensions signatures work fine (their ier implementations don't matter now):

(1) public static T2 safeGet<T1, T2>(this IDictionary<T1, T2> dict, T1 key);
(2) public static T safeGet<T1, T>(this IDictionary<string, T1> dict);

Using these, I can write

(1) var sgVal = sgDict.safeGet("b");        // ok, object { 6 }

(2) var sgVal = sgDict.safeGet<int>("b");   // ok, int { 6 }

But using this signature, a "combination" of (1)+(2):

(3) public static T1 safeGet<T1, T2, T3>(this Dictionary<T2, T3> dict, T2 key);

I caot write

(3) var sgVal = sgDict.safeGet<int>("b"); // error: "safeGet" not defined for Dictionary<string, object>

I even tried "reordering" T1-T3. I would have to write instead:

(3) var sgVal = sgDict.safeGet<int, string, object>("b");   // ok, int { 6 }

I wonder, why can the compiler correctly infer T1,T2 for (1) and T1 for (2), but not T2,T3 for (3)?