mneme-hq 0.8.0

Decision Index boundary, semantic Audit tiers, declared test evidence, and the first public decision-intent API.

This release contains all Core work since v0.7.0. Retrieval ranking, DecisionRetriever, ConflictDetector, typed-rule enforcement semantics, path applicability, and the frozen benchmark behavior are unchanged; mneme.audit/v1 is unchanged.

Added

• assess_decision_intent() (ADR-028) — the public decision-intent assessment API. External consumers (the Architecture Audit backend, integrations) ask ADR-026's intent question — is this decision text prescriptive or advisory? — through a supported, deterministic, side-effect-free boundary instead of importing private core helpers. The contract exposes only the authoritative verdict (prescriptive / advisory / neutral) with ADR-026 advisory precedence already applied; raw lexical-marker details stay private. Text-only by construction: the full text+structure intent combination stays internal to assess_protection, so intent semantics remain in exactly one place and reach consumers through versioned Mneme releases.
• Declared test evidence (ADR-024, passive) — a decision record may declare explicit test_evidence entries linking it to a test that enforces it. Ordinary mneme audit remains PASSIVE: it executes no repository-controlled code, validates declarations passively (unambiguous pytest selector, exact HEAD SHA pin, existing test file), annotates test:declared:<selector>, and never protects from a declaration alone. The VERIFIED state requires the trusted CI producer (ADR-025 design, next task).
• Published-package release smoke — after every successful Publish-to-PyPI run for a v* tag, the published bytes are validated from a clean pipx install (artifact-smoke battery), so source-checkout behavior can never be mistaken for published-package behavior.

Fixed

• Semantic Audit tier classification (ADR-026) — intent is judged from the decision text, not record shape: byte-identical decision text no longer changes tier when an unrelated structured field is added, and deterministic architectural requirements without structured fields no longer fall to Guidance. Advisory wording outranks prescriptive markers ("we should never use SQLite" stays Guidance); an installed typed rule still protects regardless of advisory prose. Identified Mneme Potential is now (Mneme-ready + Requires modelling) / protection-relevant. Human-facing mneme audit output presents Current Protection and Protection Gap (the compat-identical Identified Mneme Potential JSON field is retained).

Documentation

• ADR-023 — Canonical Decision Index and runtime-projection boundary.
• ADR-024 — Declared test-evidence ingestion (passive; shipped).
• ADR-025 — Trusted test-execution attestation (design; producer is the next planned task).
• ADR-026 — Audit tier semantics and Mneme Potential.
• ADR-027 — Decision MCP proposal-ingestion and authority boundary (documentation; no MCP implementation in this release).
• ADR-028 — Public decision-intent assessment API.

Maintenance

• Test battery policy formalized: gate / main / release / artifact-smoke / benchmark (scripts/run_test_battery.py), wired into CI and the protected-main ruleset.

Compatibility

• No changes to retrieval ranking, DecisionRetriever, ConflictDetector, existing typed-rule semantics, or the frozen enforcement benchmark.
• mneme.audit/v1 unchanged (same schema, fields; tier outcomes change for text-prescriptive decisions per ADR-026).
• CLI surface unchanged (mneme audit|setup|protect|check|benchmark).

Install

:::bash
pipx install "mneme-hq==0.8.0"