Cortex Shared-Crate Extraction

This directory is the working control plane for extracting Cortex into reusable Soma shared crates while preserving Cortex as a complete product assembled from the same libraries.

The extraction is intentionally contract-first. A folder move is not considered an extraction unless the new crate has a narrow public API, independent consumer proof, explicit dependency direction, donor-parity evidence, crate docs, and workspace architecture verification.

Donor baseline

The first reviewed donor snapshot is Cortex commit 7edf23fadb94650c2d2a2f9c80111fb44319eea8 on codex/graph-projection-lifecycle. At the start of this work that branch was two commits ahead of Cortex origin/main, so the immutable commit is the source reference for parity work.

Documents

Document	Purpose
SPEC.md	Target crate architecture, dependency graph, runtime composition, and migration sequence.
CONTRACTS.md	Rules every extracted crate and adapter must satisfy.
SOURCE-INVENTORY.md	Current Cortex modules, coupling hotspots, and planned destinations.
PROGRESS.md	Dedicated lane-by-lane extraction tracker and definition of done.
VERIFICATION.md	Required build, test, docs, architecture, parity, and smoke gates.
REVIEW.md	Review passes, findings, resolutions, and final evidence for this extraction branch.
ADR 0014	Fleet-level architectural decision governing the extraction.

Current implementation

The first proof crate is crates/shared/cortex/ingest-core (package cortex-ingest-core). It extracts Cortex's message normalization/signature logic and bounded metadata redaction. It deliberately does not know about SQLite, Axum, RMCP, Labby auth, process runtime, or deployment.

This proof establishes the pattern future Cortex crates must copy: workspace package inheritance, layer = "shared" architecture metadata, explicit features, publish = false during stabilization, README and crate-level docs, ported donor tests, and tests that exercise only the public consumer API.

Completion condition

The migration is complete only when the canonical Cortex binary is a thin composition over extracted crates, all retained Cortex behavior is covered by parity/surface tests, no shared crate depends upward into a product layer, and the obsolete monolithic source can be deleted. Until then, this tracker should make duplicated ownership and unfinished cutovers visible rather than hiding them behind a partially moved module tree.