Problem

The zeph-db crate (introduced in PRs #2371, #2372, #2379) provides a database abstraction layer, but the abstraction is incomplete and leaky:

1. 6 consumer crates hardcode sqlx with features = ["sqlite"], bypassing zeph-db entirely
2. The design uses #[cfg] type aliases instead of Rust generics (violates established project feedback)
3. Building with --features postgres (without sqlite) will fail in all consumer crates
4. zeph-memory has 367 direct sqlx:: occurrences across 31 files alongside zeph-db imports

Affected Crates

• zeph-memory (Cargo.toml:28 — hardcoded sqlx/sqlite)
• zeph-core (Cargo.toml:50 — hardcoded sqlx/sqlite)
• zeph-orchestration (Cargo.toml:19 — hardcoded sqlx/sqlite)
• zeph-index (Cargo.toml:23 — hardcoded sqlx/sqlite)
• zeph-mcp (Cargo.toml:35 — hardcoded sqlx/sqlite)
• zeph-scheduler (Cargo.toml:23 — hardcoded sqlx/sqlite)

Implementation Phases

• Phase 1: Remove direct sqlx deps from consumer crates, re-export from zeph-db, add feature propagation
• Phase 2: Migrate from #[cfg] type aliases to Pool<DB> generics per feedback_db_generics.md
• Phase 3: Port test infrastructure, add postgres CI matrix

Architectural Review

Full report: .local/handoff/db-arch-review.md

Key findings:

• Qdrant placement in zeph-memory is CORRECT (do not move to zeph-db)
• VectorStore trait abstraction is well-designed
• Dialect trait and FTS helpers are solid
• sql! macro postgres variant has a memory leak (Box::leak without LazyLock caching)
• Feature flags: postgres correctly excluded from full, mutual exclusion enforced
• Pattern compliance: thiserror, no async-trait, rustls, no hardcoded models — all pass