LT Spice like almost all computer circuit analysis systems uses a nodal approach - the unknowns are voltages at each node, KCL gives the sum of currents into a node as zero. Impedances then give a matrix relationship between the nodal voltages and the contribution to the nodal currents, e.g. a resistor between nodes 1 and 2 would give the matrix relationship:
I1 = 1/R -1/R . V1
I2 -1/R 1/R V2
(difficult to format matrices). All the contributions can be added for a linear circuit to give an overall G matrix which relates external currents as a vector to an overall V vector
I = G.V
then G is essentially inverted (in SPICE LU decomposition is used as the process needs to be repeated many times for each linearisation of the nonlinear circuit).
So LTSPICE thinks in terms of currents into nodes. For hand analysis it is easier in this example to think in terms of loop currents (I1 and I2) and just use KVL to state the
sum of voltages around a loop is zero (instead of the sum of currents into a node is zero). I3 is neither a nodal current nor a loop current so is best looked at as short hand
for the sum of I1 + I2.
Sorry, a bit long winded, but what I'm trying to say is that if you use LTSPICE to check results you need to convert nodal currents back into loop currents and I would guess it
is quite easy for signs to misbehave. Especially given that I3 doesn't really exist as a physical entity separate from I1 and I2.