MAINT: Raise test coverage on _arma, _lss, _kalman

[mmcky]

Problem

Package coverage is strong overall, but three long-standing core modules lag well behind (figures from the current instrument; re-baseline after the coverage-measurement fix in this milestone lands):

Module	Coverage	Untested regions
_arma.py	53%	property setters (phi/theta), simulation/impulse-response paths
_lss.py	75%	validation branches, simulate_linear_model region, stationarity paths
_kalman.py	80%	stationary_* computations, smoothing branches

Notably, simulate_linear_model is one of the six names missing from the top-level namespace (#853) — untested code and broken export travel together.

Proposed change

Targeted tests, not ceremony: setter round-trips and seeded simulation shape/determinism checks for _arma; constructor-validation pytest.raises cases and a simulate_linear_model statistical sanity check for _lss; the stationary-solution fixed-point property (residual of Σ = A Σ A' + C C') for _kalman. Aim ≥85% per module; don't chase plotting/__repr__ lines.

Acceptance criteria

• Each module ≥85% against the post-fix coverage baseline
• New tests are deterministic (seeded) and fast

---

From the July 2026 technical-debt audit (AI-assisted; claims verified against 28d4b3b on 2026-07-25).