states.KalmanUpdate

Reads predicted mean ``cs.{field}`` of shape ``(N, D)`` and covariance ``cs.{cov_field}`` of shape ``(N, D, D)``, reads detection measurements ``ds.{field}`` of shape ``(M, M_dim)``, and writes back the updated mean and covariance for matched tracklets. Unmatched tracklets keep their predicted state (predict-only on miss). The update applies the standard Kalman gain ``K = P H^T (H P H^T + R)^{-1}`` via ``solve_psd``, then propagates covariance through the Joseph form ``(I - K H) P (I - K H)^T + K R K^T`` to preserve positive semi-definiteness under finite-precision arithmetic. The result is symmetrised to remove off-diagonal drift that accumulates over long sequences. ``H`` shape ``(M_dim, D)`` is the measurement matrix and ``R`` shape ``(M_dim, M_dim)`` is the measurement-noise covariance. Each bbox/centroid process ships a ``make_update(...)`` factory that constructs the matching :class:`KalmanUpdate`.