Status: shipped in the current development line. The workspace catalog has a single lexical backend boundary and a single semantic vector authority.

Code owns file admission, deterministic chunking, stable chunk IDs, source digests, revision fencing, path filters, result projection, and current-source verification. The backend only indexes the admitted chunk text and returns bounded ranked ordinals.

The product default is the pure-Rust a3s-vec FTS engine, identified in result metadata as a3s_vec_fts_v1. Every partition, including a single-document partition, goes through that engine so production behavior is the same at every workspace size. A deliberately minimal build can select the dependency-free scorer, identified as portable_bm25_v1; that is an explicit build-time fallback and does not silently change the configured engine.

Semantic retrieval uses the exact a3s-memory in-memory adapter. Embeddings are validated once, grouped into bounded provider requests, and published as immutable per-file partitions. There is no second semantic index, migration shadow, or runtime authority selector.

For each admitted catalog partition, the a3s-vec adapter:

1. Reuses Code's canonical tokenizer and normalizes the tokens into one FTS field. This keeps identifier splitting and CJK bigrams identical across exact, lexical, and hybrid channels.
2. Creates a temporary collection with a whitespace FTS index and inserts documents in bounded batches. Engine primary keys are generated from dense ordinals; the original Code chunk ID stays in the surrounding catalog.
3. Flushes and closes the collection before publishing the partition. Queries reopen read-only for the shortest possible scope, map keys back to immutable chunks, and discard non-finite scores.
4. Retains only the persisted temporary directory and compact key/term metadata. The source catalog remains authoritative and owns cleanup.

The current catalog replaces files atomically, so a per-file collection is intentionally ephemeral. At most four hot read-only collections are retained across sessions; colder partitions open, query, and close within one bounded operation. This keeps reopen cost bounded while preserving immutable snapshot semantics.

The indexed lexical path first uses term metadata to select candidate files, then asks a3s-vec for bounded top-k FTS results per selected partition. Exact literal, optional Code Intelligence symbols, lexical, and semantic channels are fused with deterministic reciprocal-rank fusion (k=60). The optional CPU reranker examines a bounded candidate pool, protects exact identifiers, and applies overlap/boilerplate diversity. It returns the original RRF order when configuration or scratch budgets are invalid.

The managed grep path remains exhaustive for literals and regular expressions. It does not depend on an embedding provider or durable index. Every semantic or lexical snippet is reread through WorkspaceServices and must match its recorded digest and byte range before it reaches the model.

a3s-vec is a pure-Rust crate. Product builds enable the a3s-vec-fts feature; no native sidecar library, ZVEC_LIB_DIR, or platform packaging script is required. Unsupported or minimal embeds can compile with --no-default-features and use the portable scorer.

Persistent generations use an atomic CURRENT pointer and a schema-v2 manifest. On reopen, Code verifies every retained chunk's text digest, stable chunk identity, byte/line range, canonical source digest, and duplicate-ID constraint before exposing the collection. A damaged or incompatible generation (including prior a3s_vec_fts_v1 generations) is treated as absent and is rebuilt from the current catalog; it is never served as a partially trusted index.

• The locked exact/identifier/CJK relevance fixture passes for both engines.
• Catalog creation, replacement, deletion, lag recovery, budget rejection, stale-source filtering, and cleanup are atomic and deterministic.
• Concurrent replacement and query tests pass without resource exhaustion.
• Default and --no-default-features builds compile and run their respective lexical suites; explicit a3s-vec selection fails with a typed configuration error when the feature is absent.
• Node.js, Python, and Go expose the same lexical enum, status/result shape, typed chunking/reranking options, cancellation, and close semantics.

Rollback is configuration-only: disable semantic/hybrid modes or select the portable lexical engine in a minimal build. Exact search, managed grep, Code Intelligence, and Agent memory stay available.