Highlights
What this release added
- Added an off-by-default, provider-independent decision-policy boundary across chat, Today/Journey, voice check-ins, and connector quest automation while preserving every existing production choice.
- Added privacy-safe candidate, feature, decision, propensity, and outcome contracts plus a reproducible candidate-level dataset exporter and shadow evidence report.
Fixes
What this release corrected
- Restored compatibility for long-lived databases whose legacy notification queue columns remained required and whose activity-log ID lost its generated default; quest audits now also supply their UUID explicitly, and database verification enforces both runtime insert contracts.
- Kept chat decision-policy `shown` linkage aligned with the cards actually rendered by excluding hidden template fallbacks whenever a live chat, connector, or LLM suggestion is available; the complete candidate set and existing production-selected candidate remain unchanged.
- Made personalization RAG capability-aware and serverless-safe: Vercel uses the existing TypeScript/context fallback without importing Python, rebuilding FAISS indices, or writing under the application directory; explicit local Python keeps its cache under the OS temporary root.
- Prevented immediate recommendation outcomes from racing and being lost before their matching decision/candidate transaction becomes visible, without blocking the user action.
- Replaced the decision ledger's unsupported Neon HTTP `db.transaction()` call with the driver's atomic batch API after production off-mode smoke exposed zero persisted decisions.
- Made decision-policy dataset and evidence commands select local, preview, or production data explicitly and report source SHA, observation start, mode, linkage, and runtime privacy coverage.
Known issues
What remained constrained
- The production evidence contains 33 off-mode diagnostic decisions from one controlled user group. These pre-shadow rows are ineligible for v0.54 training and do not support policy-superiority, causal-lift, or counterfactual-reward claims.
- No trained decision policy exists, shadow scoring remains inactive, and the existing production selector remains authoritative. A separate approved activation must establish the future evidence observation window.
- Redis is operating in an explicitly approved degraded mode through 2026-08-01: stale TCP and REST endpoints are removed, DNS error noise is closed, and affected cache/anomaly capabilities use their documented bypasses until a valid supported Upstash resource is restored.