ADR 0001: Incremental repo-state embeddings (diff-based)

Date: 2026-01-04 Commits: a00c179

Context: Repo-state ingest enumerates the full git tree at each commit and performs per-file filtering; on large repositories this can dominate scheduler throughput. Decision: Persist a per-commit repo-state aggregate (sum vector + eligible file count) and derive child aggregates by applying parent..child diffs (add/modify/delete/rename) using GitPython, embedding only missing blobs via the existing file cache. Constraints: Fall back to a full recompute when root ignore files change (.gitignore, .loreleyignore), when the commit has multiple parents (merges), or when diff parsing fails, to preserve correctness. Consequences: Most ingests become O(changed_files) and avoid full-tree traversal and full-cache reads; merges and commits that invalidate incremental derivation still require full recomputation.