Structured covariance prerequisite

Implement the existing shared receiver-error family as an innovations filter: white three-axis error, up to two exponentially correlated components, one constant offset and one full-record-normalized drift component. Preserve all cross-axis covariance, actual timestamps and declared coordinate transforms.

The implementation must match the existing dense reference for whitening, log determinant, inverse whitening and coordinate-subset marginalization. Residuals and every Jacobian column use the same operator. Do not construct whole-record covariance, Cholesky or projection matrices.

Keep block continuation explicit: covariance configuration, ordering, direction and column count must match; missing state cannot silently reset a later block. Use positive-definite innovation solves without jitter, eigenvalue clipping or dropped nuisance columns. Treat covariance as supplied, not observationally estimated or calibrated. No instrument bounds or Earth-model results change.

Qualify on small independent dense comparisons, including irregular timing, cross-axis terms, singular latent components, different coordinate units and retained-axis permutations. A later separately frozen finite benchmark will measure memory/runtime at long-record scale; passing small tests is not that benchmark or qualification of native streaming integration.

This prerequisite supports training-only covariance development and ESA-scale prediction. Bounded calibration/covariance-estimation uncertainty and predictive adequacy remain separate requirements. No archive fits or protected ESA outcomes are evaluated by these software controls.