v2.50.37

Local embedding recall now binds the query/document input profile and vector dimension to the stored index. Explicit Nomic task prefixes and EmbeddingGemma fact-checking profiles are available on refresh, query and recall evaluation. Profile or dimension changes require re-embedding; legacy vectors retain raw semantics.

Embedding calls now retain durable local attempt receipts and current owned-local proof. Malformed, zero, nonfinite, mixed or mismatched vectors refuse rather than producing scores. No automatic model download, metered fallback or new spend authority is added.

A frozen 40-query/172-document SciFact subset compared raw/prefixed Nomic and Gemma768/Gemma256. The 108 local attempts reconciled to 324 receipts at $0; an offline replay reproduced the metrics. Nomic prefixes did not improve this fixture. Raw remains the default; these subset results do not certify general retrieval quality, expertise or S0. See the retained comparison record.

Clean-wheel MCP conformance for both supported protocol eras, host checks and six installed local CLI checks passed before publication. Canonical fixture beliefs remained unchanged.

Full offline unit suite: 13,215 passed, 20 existing skips and 85.54% combined coverage. The independent coverage gate and all local source checks passed.

CI on exact main efdb70210d1427499c4e389e4c5f1263e684b9f3 passed before publication, including Python 3.12, 3.13 and 3.14.

The public release installer passed in isolated manager and data roots. Installed CLI reports 2.50.37; offline doctor, JSON inspection, noninteractive setup refusal and read-only decision-review checks passed. All three embedding commands expose profile/dimension options. Synthetic offline embedding checks retained distinct role inputs and complete $0 lifecycle receipts.