Self-contained type definitions for the new storage layer replacing
Graphology + JSON. Covers all 6 graphs, cross-graph links, tags,
attachments, projects, team, and workspace metadata.

Key design decisions:
- All entity IDs numeric (autoincrement), slugs for external use
- Store is fully synchronous (better-sqlite3), embeddings passed from outside
- SearchQuery supports text (FTS5), embedding (sqlite-vec), or both
- Cross-graph links via junction table, no proxy nodes
- Tags and attachments as shared stores across all graphs
- ProjectScopedStore for multi-project workspaces
- TeamStore at workspace level with numeric member IDs
- MetaMixin on all stores for arbitrary key-value metadata
- Detail types (NoteDetail, TaskDetail, etc.) include relations + crossLinks

Comprehensive proof-of-concept validating SQLite stack for store:
- sqlite-vec v0.1.9 loads with better-sqlite3
- WAL mode, foreign keys, CASCADE deletes
- FTS5 keyword search with sync triggers
- Vector KNN search (384d + 1024d)
- Hybrid RRF fusion (FTS5 + vector)
- JSON arrays via json_each()
- Nested transactions (savepoints) + rollback
- PRAGMA user_version migration pattern
- File persistence (close/reopen)

Benchmarks (384d/100k): insert 28µs/row, vec 36ms/search, FTS5 0.5ms/search
Benchmarks (1024d/100k): insert 47µs/row, vec 99ms/search, FTS5 0.6ms/search

Key findings:
- sqlite-vec requires BigInt for rowid (db.defaultSafeIntegers(true))
- vec0 not cleaned by CASCADE — store must DELETE manually
- 384d recommended over 1024d for large datasets (3.6x faster search)

…ed schema

Major refactoring of store types and schema:
- Unified Edge type replaces Relation + CrossLink (one edges table)
- Code, Docs, Files: single table per graph with kind field (no separate file/symbol tables)
- Graph-prefixed table names: knowledge, tasks, epics, skills, code, docs, files
- Hierarchy via edges (file→symbol, file→chunk, dir→file) instead of FK parent refs
- Detail types simplified: edges: Edge[] instead of separate relation arrays
- SearchResult.id changed to number (numeric IDs throughout)
- All 22 tests passing

…ed schema

- Unified Edge type replaces Relation + CrossLink (one edges table)
- Code, Docs, Files: single table per graph with kind field
- Cleanup triggers on all entity tables (edges + attachments + vec0)
- Tags as entities with edges (kind='tagged'), inline in records
- Attachments: metadata only (url optional, no BLOB)
- CASCADE on all project_id FKs
- Search tests: keyword, vector, hybrid, project isolation
- All imports use @/ alias
- 31 tests passing

…ests

- SqliteTeamStore: CRUD, slug uniqueness, getBySlug, list, MetaMixin
- SqliteProjectsStore: CRUD, cascade delete via FK + triggers
- SqliteAttachmentsStore: metadata registry (add/remove/list), optional url
- Wired team + projects in SqliteStore
- Fixed cleanup triggers: code/docs/files now also clean attachments
- Cascade delete tests: edges, attachments, vec0, tags all verified
- 70 tests passing (31 sqlite-specific + 39 contract)

Knowledge, Tasks (with Epics), Skills stores with UUID slugs, version
conflicts, hybrid search, tags via edges, batch fetching, and cleanup
triggers. EntityHelpers extracts shared tag/attachment/edge logic.

Indexer-driven stores with updateFile/removeFile pattern, cleanup via
triggers, hybrid search (FTS5+vec0 for code/docs, LIKE+vec0 for files),
cross-file edge resolution, directory auto-management, and json_each
symbol lookup for docs.

… fixes

ProjectScopedStore wraps all sub-stores with fixed projectId, cached
via store.project(id). Full edge CRUD at both scoped and workspace
levels. GraphName extended with epics/tags.

Audit fixes: recursive ensureDirectory chain, minScore in all
FilesStore search modes, orphaned tags cleanup in setTags, removed
template literal SQL interpolation, simplified isExported check.