S7 — SLAM-Feature Update (optional)

An optional layer above the MSCKF core: instead of marginalizing every feature after it leaves the window, keep long-lived landmarks in the state and update them by reprojection. Persistent landmarks re-observed over long spans bound drift better than the pure sliding-window MSCKF.

What it does

Features are kept in-state (3 parameters each, or 1 with an anchored inverse-depth parameterization) and updated by reprojection without marginalization. A long-lived landmark seen across many keyframes acts as a persistent constraint.

The contract

Signature	features kept in-state (3 or 1 params each), updated by reprojection, not marginalized
Post / invariants	in-state feature P blocks stay PSD and correlated with the poses that see them · feature initialization into the state uses stochastic init (correct cross-covariance — the same contract as S3) · pruning a SLAM feature is a Schur complement, info-preserving
cortex	features/representations.hpp defines the parameterizations; the SLAM-feature update is not wired — the running filter is pure-MSCKF

Status — optional, not implemented

The cortex filter is pure-MSCKF: it marginalizes every feature (S6) and keeps no persistent landmarks in-state. The inverse-depth parameterizations exist (features/representations.hpp), but the in-state SLAM update is not wired. This stage is documented for pipeline completeness and is not relevant to the #212 over-confidence (which is a pure-MSCKF consistency problem). It becomes relevant when persistent landmarks are added to bound long-horizon drift.