lti_system_gen delayed-causal state assembly is live but unverified and produces marginally-stable systems

[dantzert]

On main, lti_system_gen builds a global A/B/C by splicing each input's lti_from_gamma LTI subsystem into the state-space for delayed ("d") forcing edges. The block that does this is flagged in lti.py:

### everything below this point is garbage. not performing at all as desired at the moment

This is not dead/commented-out code — it is the live path that runs whenever an output has a delayed-forcing edge. Verified empirically on the documented known_topology test fixture (x2←delayed-u1, x8←immediate-u2, x9←immediate-x8, x9←delayed-x2) with bibo_stable=True:

• assembled A is 103×103 (state explosion from the gamma-subsystem expansion),
• 2 eigenvalues have real part ≈ 0 (max = 0.0) → the system is marginally stable / has an integrator mode, NOT Hurwitz — even though bibo_stable=True is supposed to guarantee stability,
• the only test (test_lti_system_gen_returns_state_space in tests/test_modpods.py) checks return-shape + simulatability, not correctness of A/B/C.

Stale TODOs remain: lti.py:344 ("incorporate bibo stability constraint into instantaneous fits"), lti.py:644 ("cast strings back to tuples" for the swmm path).

This is the missing bridge between #16 (LTI impulse-response kernels) and the stated promise of an "LTI state-space model suitable for control design."

Proposal: add a correctness test (assembled A/B/C reconstruct a known cascade / are Hurwitz for bibo_stable=True), then either fix the splicing or remove the broken expansion.

• Cost: High (rewrite the matrix-splicing logic + a correctness test).
• Benefit: Very high (delivers the control-design use case end-to-end; makes Support other convolution kernels  #16 useful).