Compatibility notes for the October 2026 audit fixes¶

Model identity¶

Managed local models now identify the effective embedding function: actual pack or model-file bytes, tokenizer/configuration and effective preprocessing, pooling, prefixes, truncation, normalization and task behavior. EmbeddingSpace.revision contains the versioned fingerprint; its public shape stays the same. An explicit external embedding space remains caller-owned. HTTP and MCP derive their space from an installed model instead of assigning a nominal label.

A managed store with an old nominal identity, or one whose effective model changed, fails compatibility checks before new writes. Back up the store and export its retained source text, IDs, versions, metadata and URIs with the compatible earlier binary. Create a separate store using the intended model, regenerate embeddings, and compare counts, source hashes and query fixtures before switching paths. When text was not retained, use the original sources. Do not edit store.json to relabel vectors produced by a different function.

Imported identities and checkpoints¶

Importers retain v1 destination IDs by default. They now stamp a typed source identity and refuse a conflicting destination owner, including ambiguous v1 namespace/part IDs. Use --id-mapping v2 for unambiguous typed hash IDs. Mapping-version changes can create a second set of destination IDs; choose a fresh destination or explicitly reconcile the old records.

Existing unstamped records require --adopt-legacy and an exact source/namespace/source-ID/ part-layout match. An ID match alone does not establish ownership. Changed split layouts fail by default; --split-policy replace writes the complete replacement before conditionally deleting obsolete owned parts. Checkpoints advance after that cleanup and verify the committed source content on resume. This is migration/resume support, not general source-deletion sync.

Retrieval and resource errors¶

Normal sparse retrieval consistently reports Jaccard similarity; Hopfield attention remains an explicit separate operation. Exact dense selection uses canonical f64 scores. Approximate indexes still have candidate-recall tradeoffs; reranking does not make their recall exact.

Rust reasoning callers should use multi_hop_with_limits and synthesize_concepts_with_limits. Their legacy Vec wrappers panic on budget exhaustion. Node propagates checked failures. Filtered SDK queries select eligible IDs before ranking. Exact cosine SDK paths merge bounded ID groups; fused/model-text paths reject oversized selections when their scores cannot be merged correctly. Reduce the selection or use a suitable server-side search contract.

Persistence and format¶

New stores use schema 5 for scoped meaning maintenance. Existing schema 2/3/4 stores retain their schema during ordinary opens and writes; durable governance returns explicit upgrade guidance. Upgrade a copy with warm-format plan /path/to/store, then warm-format upgrade /path/to/store /path/to/upgraded-store. Verify the copy before switching paths. The upgrade preserves source bytes and existing checkpoints; it does not reembed or relabel model vectors. Older binaries reject schema 5 instead of omitting its scoped mutations. Node offers governMemoryAsync() for maintenance off the event loop; the synchronous governMemory() remains available and reports checked errors.

Failed WASM saves retain the in-memory store for retry. Mobile snapshots reject destinations within the live store after canonical path resolution. The server restores a tenant through a recoverable intent rather than relying on in-process rename rollback.

The original audit is a historical baseline. Its scores do not describe the fixed implementation; remediation evidence records what was implemented and validated, with remaining verification limits.

Embedding limits and text encoding¶

Local embedding calls now enforce positive configurable text, byte, token, output and concurrency limits; split larger calls or configure EmbedOptions::limits for the workload. The soft batch token budget remains separate. Embedding reports expose both truncation stages. Rust model callers should use try_encode_text and try_find_analogy to handle failures; the legacy infallible wrappers panic instead of turning failures into valid empty vectors. Durable writes and Node APIs propagate checked errors before mutation.

The default text encoder remains V1. Opt into multiscale-words-v2 by reencoding original sources into a new store; do not relabel an existing schema or rewrite old golden vectors.