[mdabydeen]

Closes #3.

What

examples/radar_tracking_ukf.rs printed a position-error column that grew from 1.47 m to
26.84 m over 15 steps, then closed with "UKF tracked continuous curved flight without
analytical Jacobian matrices." The number and the conclusion contradicted each other.

Root cause

• The transition closure held y constant (|state, _d| [state[0] + 8.0 * _d, state[1]])
  while the ground-truth y moved, so the model and reality did not agree.
• A single range observation sqrt(x^2 + y^2) is barely observable in y, so the estimate
  stuck.
• The "tracked" line was unconditional, with no check on the error.

The filter math was fine; the demo model was inconsistent.

Fix

• Model y in the transition: the target now flies a constant-velocity straight line
  (x0 + vx*t, y0 + vy*t), which is what a linear transition can actually track.
• Add a second, observable sensor (a lateral y reading) so the 2-D state is recoverable.
• Report success honestly: print an error-qualified "tracked" line only when the final
  error is under a threshold, guarded by an assertion so a future regression panics the
  example instead of silently printing a false success.

Verification

• Final error now stabilises at ~0.3 m; the demo prints an honest, error-qualified line.
• Regression guard confirmed to fire: temporarily tightening the threshold to 0.05 m made the
  example panic (exit 134), then reverted.
• cargo fmt -- --check clean
• cargo clippy --all-targets --all-features -- -D warnings clean (examples included)
• cargo run --example radar_tracking_ukf exits 0; the other three examples still run

Note: examples are not part of cargo test, so correctness is guarded by the in-main
assertion plus the before/after numbers, not a unit test.