ELIPS v1.0 is a vertically complete embedded engine. The list of deferred items is intentional: each future capability has a v1.0 hook so it can land additively, without breaking the surface users build against today.

Shipped in v1.0

C++23 core with hexagonal layering and explicit C++ Core Guidelines alignment.
Metrics — cosine, euclidean, dot product — with NEON SIMD plus scalar dispatch.
Indexes — HierarchicalGraphIndex (HNSW) and ExactIndex — behind a single IndexPort.
Durable storage: IDENTITY, atomic snapshot, record-based WAL with CRC32C, deterministic crash recovery.
Single-writer advisory file locking; shared read-only opens for multi-reader serving.
Document-aware records — DocumentAttachment, ChunkInfo, EmbeddingLineage — persisted through WAL::insert_ex.
Native place_document, seek_text, seek_hybrid, explain_seek.
Built-in local text embedder with automatic default provisioning for new databases.
MetadataIndex acceleration for equality and set-membership filters.
Segmented persistence with elips.manifest plus per-vault segment files; compact() rebuilds indexes and rewrites segments.
Dynamic metadata and the Filter predicate engine, shared by the SDK builders and the EQL parser.
Atomic batched transactions with eager validation and RAII rollback.
EQL — lexer, recursive-descent parser, AST, executor — covering seek / fetch / scan / place / erase.
elips CLI; PyBind11 Python bindings with py.typed; benchmark suite.
Optional GPU index family behind GpuPort with a documented backend fallback chain.

Deferred (with v1.0 hooks)

Future capability	v1.0 hook
Per-segment indexes + compaction	IndexPort; snapshot/WAL split per vault
Full MVCC version chains / Snapshot Isolation	Single-writer model + transaction buffer
Quantized indexes (PQ / OPQ / SQ), DiskANN	IndexPort + make_index factory
AVX2 / AVX-512 distance kernels	Function-pointer dispatch seam in Metrics
Columnar metadata, attribute B-trees, inverted / bloom	Filter over Payload today
Multi-reader shared locks beyond current model	LockManager seam
Multi-node replication / sharding	WAL is a logical, streamable log
Cloud object-storage adapters (S3 / GCS / Azure)	StoragePort adapter pattern
NumPy zero-copy ingestion, async/streaming C++ APIs	Binding + SDK extension points
Cross-platform snapshot encoding (little-endian)	Centralized Serialization

These items are not in v1.0 today. Do not treat them as available behaviour; they are recorded here so the design seams that make them additive stay visible to contributors.

Known v1.0 limitations

Snapshot serialization uses native byte order — single-machine use only.
Checkpoints rewrite the whole snapshot (O(N)).
Filtered ANN search post-filters an over-fetched candidate set rather than expanding ef adaptively.

See design decisions for the ADRs that capture why these trade-offs were made.