Fix input-derivative pairing; default to inline linear-SCC reassembly

[baggepinnen]

Two changes to the linearization pipeline, found while linearizing a planar two-joint robot (a model whose large linear SCCs take the numerical LinearSolve-based path under the default reassembly).

Fix: input-derivative column pairing in causal_simplification call sites

ModelingToolkit.linearize with allow_input_derivatives = true returns B with 2nu columns, where column nu + i is the derivative of input i. The pairing passed to causal_simplification was derived by broadcasting 1:nu against the set of nonzero derivative columns:

duinds = findall(any(!iszero, eachcol(matrices.B[:, nu+1:end]))) .+ nu
u2du = (1:nu) .=> duinds

When exactly one derivative column is nonzero this silently broadcasts to [1 => k, 2 => k, ...], pairing every input to the same derivative channel; other partial subsets throw a DimensionMismatch. The result is a wrong reduced system whenever inputs enter retained algebraic equations. On the robot, the input complementary sensitivity collapsed to a rank-1 projector (Ti(0) ≈ [1 1; 0 0] instead of I), i.e. the loop transfer's input columns came out identical.

The pairing is now positional over all inputs, [i => i + nu for i in 1:nu]; all-zero derivative columns contribute nothing (and filtering them out is not an option, since causal_simplification infers the input count from the number of pairs).

Verified on a minimal 5-equation model with a 3×3 linear SCC and two inputs: the numerically-reassembled linearization now matches the symbolically-inlined one to a relative error of 6e-16 (previously ~100% error); the robot's input complementary sensitivity matches to 8e-13 across frequency.

Default: inline linear-SCC reassembly for all linearizations

DEFAULT_LINEARIZE_KWARGS = (; reassemble_alg = DefaultReassembleAlgorithm(; inline_linear_sccs = true, analytical_linear_scc_limit = 1)) is now forwarded by every linearization-based function (named_ss, get_named_sensitivity/get_named_comp_sensitivity/get_named_looptransfer, batch_ss/batch_linearize, trajectory_ss, build_quadratic_cost_matrix). Caller-supplied keyword arguments take precedence, so the ModelingToolkit default remains reachable.

With the pairing fix above, both reassembly paths produce correct linearizations; the inline default is kept because it avoids the input-derivative-augmented form entirely (no spurious marginal modes, better conditioned realizations) on models with linear SCCs.

🤖 Generated with Claude Code