8:50 AM
First of all we need to figure out how to compute the Levi-Civita connection of a metric in any manifold, regardless whether the elements of it are thought as vectors, functions, polynomials or matrices. In the numerical computations such a manifold will be finite-dimensional, that is every time we work with a presentation of this manifold as a piece of R^n with all the structures pulled back to this piece via the chart.
The second fundamental form is then the normal part of this Levi-Civita connection applied to two vectors which are tangent to your submanifold, i.e. the curve in your particular problem.
The Koszul formula is the explicit expression that gives the Levi-Civita connection in terms of the metric and the Lie bracket:
<\nabla_X Y, Z> = 1/2 (X g(Y,Z) + Y g(Z,X) - Z g(X,Y) + g([X,Y],z) - g([Y,Z],X) - g([X,Z],Y))
Using this formula we obtain the usual expression for the Christoffel symbols in a coordinate frame, because the brackets of partials vanish. If the frame is arbitrary the brackets produce the "commutation coefficients", and we get the generalized expression for the Christoffekl symbols in a (possibly) non-holonomic frame.
With this in mind there should be a way to pick a concrete example of a manifold and a local frame in an open set of this manifold, and compute the connection coefficients of this frame. This should not involve any metric or connection!
The local frame that we are looking at the moment should be defined locally, that is on an open set, not just along a particular curve. In this situation, using the coordinates (!!!) we should be able to calculate directional derivatives using the standard multivariable calculus formulae.
These coordinates come from the chart in which we represent all the objects that we are dealing with, including the vector fields that form the local frame
Finally, when you endow the curve with a convenient local frame there should be a natural (for you choice) way of extending this frame to a tubular neighborhood of the curve. The theory then must assure that the constructions that you use do not depend on this choice.
Therefore, if all the analytic expressions that are in the foundation of your algorithms are correct, you should get consistent results, but there may also be issues with stability etc.