fix(db): restore postgres backend compile and sqlite/postgres parity

[bug-ops]

Summary

Restores the PostgreSQL backend and the SQLite/PostgreSQL schema parity for the data layer. Three closely related data-layer correctness defects, fixed together:

• zeph-memory agent_sessions: SqliteStore fails to compile when postgres feature is also enabled (--all-features) #4964 — PostgreSQL backend did not compile. SqliteStore::upsert_agent_session bound the turns count as a bare u32, which only implements sqlx::Type<Sqlite> (PostgreSQL has no unsigned integer types). This pinned the query to the SQLite driver and broke the PostgreSQL backend with E0271 ("expected Sqlite, found Postgres") on any build that activates the postgres feature. turns is now bound via cast_signed() (i32, matching the INTEGER column) and the read tuple decodes it as i32 so PostgreSQL does not reject the INTEGER/INT4 column as an i64 mismatch. A full-workspace cargo check --no-default-features --features postgres now compiles cleanly — this was the only PostgreSQL-breaking bind site in the workspace.

• zeph-db pool.rs: connect() body empty without sqlite/postgres features causes compile error #4956 — consumer crates could not be built/doc-checked in isolation. Crates depending on zeph-memory/zeph-db without selecting a backend (zeph-sanitizer, zeph-acp, zeph-bench, zeph-tui, and zeph-skills via its optional qdrant dependency) tripped the zeph-db compile_error! plus a follow-on cascade, because the default-features = false workspace deps stripped the default backend. Each crate now carries the standard default = ["sqlite"] / sqlite / postgres feature trio forwarding to zeph-memory, matching the established zeph-mcp/zeph-index pattern. The backend-only imports in zeph-db/src/migrate.rs are gated behind any(feature = "sqlite", feature = "postgres") so a genuine no-backend build surfaces only the clear compile_error!.

• zeph-db sqlite/postgres migration sequences diverge (file count + content parity) #4957 — SQLite/PostgreSQL migration sequences diverged (P2, system-invariant 001 §13). PostgreSQL was missing the fact_access_log table and messages.memory_tier/messages.qdrant_promoted columns (five-signal SYNAPSE retrieval, feat(memory): extend SYNAPSE to five-signal retrieval + async consolidation daemon #4374) and the implicit_conflict_candidates table (graph conflict detection) — present only in SQLite, so PostgreSQL deployments ran a divergent schema. Added forward-only PostgreSQL migrations 101_five_signal_retrieval.sql and 102_implicit_conflict_candidates.sql (integer columns sized BIGINT to match the i64 Rust accessors), plus a SQLite parity placeholder 101_trajectory_memory_cascade.sql (the cascade is inline on SQLite but needs a dedicated drop-and-re-add migration on PostgreSQL) to keep file counts equal (101 = 101). A new migration_parity integration test fails the build if the dialects ever diverge in file count, logical migration name, or defined table set.

Verification

• cargo +nightly fmt --check — clean
• cargo clippy --workspace --all-targets -- -D warnings — clean
• cargo nextest run --workspace --lib --bins — 10555 passed, 0 failed, 21 skipped
• cargo test -p zeph-db --test migration_parity — 3 passed (file-count / logical-name / table-set parity)
• cargo nextest run -p zeph-memory -E 'test(agent_session)' — 8 passed
• cargo check --no-default-features --features postgres — clean (PostgreSQL backend compiles end-to-end)
• cargo check -p {zeph-sanitizer,zeph-acp,zeph-bench,zeph-tui} --lib and -p zeph-skills --lib --features qdrant — all clean (previously failed)
• RUSTFLAGS="-D warnings" cargo check --workspace --all-targets --features desktop,ide,server,chat,pdf,scheduler --locked — clean
• RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --all-features -p zeph-db -p zeph-memory — clean (one pre-existing private-intra-doc warning, not introduced here)
• cargo test --doc -p zeph-db -p zeph-memory — 39 passed

Notes / follow-ups

• A live PostgreSQL round-trip (e.g. binding String to the agent_sessions TIMESTAMPTZ columns) is a separate, runtime-only concern not exercised by any current CI job; tracked as a follow-up.
• The de-facto backend model is "SQLite always present, PostgreSQL wins by priority when both features are on"; this contradicts the literal wording of system-invariant 001 §13 ("enabling both is a compile_error!"). The code only errors on no backend. The invariant text should be reconciled with the implemented behavior in a follow-up.

Closes #4964
Closes #4956
Closes #4957