fix: Tier 1 ship-blockers — packaging, pg value integrity, eager delete (v1.0.1)

[revxshafi]

Summary

Tier 1 correctness fixes. No public API changes and no breaking changes — safe as a patch release (1.0.0 → 1.0.1).

• NEW-13 — import 'sql-switch' no longer drags in both drivers. With tsup splitting: false, esbuild inlined each lazily-imported driver into the entry chunk and hoisted its top-level import pg / import better-sqlite3, so a single-engine install crashed on import. splitting: true keeps the drivers as their own chunks, loaded only once an engine is selected — restoring the optional-peer-dep invariant. Covered on both the ESM and CJS entrypoints.
• NEW-2 — pg get() no longer corrupts JSON-looking strings. get() read through drizzle's jsonb column, running a second JSON.parse on a value pg had already parsed, so a snowflake id "123456789012345678" came back as a precision-lost number and disagreed with entries()/scans on the same row. get/set/delete now use the same raw parameterized $n::jsonb pool path as the bulk upsert and the scans, so every read/write path agrees byte-for-byte.
• NEW-7 — pg set(null) is consistent across paths. drizzle mapped a JS null onto SQL NULL, which the NOT NULL value column rejected on the forced path. set now stores a jsonb null on every path; get() reads it back as null and has() still reports the row.
• C1 — an un-awaited delete() now lands. delete() only did its work inside the WriteOperation callbacks, so a fire-and-forget delete() silently did nothing while a fire-and-forget set() committed. It now executes eagerly at call time (matching set()) and evicts any buffered set for the key so a later flush can't resurrect it; await/.force() only decide whether you wait.

drizzle-orm is still used for the SQLite driver and the schema builders, so it remains a dependency. package.json peer/optional deps are unchanged — no driver was promoted to a hard dependency.

Tests added

• test/no-driver-hoist.test.ts — dist guard: asserts neither built entrypoint (ESM + CJS) has a module-scope import/require of pg/better-sqlite3. Proven to fail against a splitting: false build.
• test/eager-delete.test.ts — engine-agnostic (KeyProxy over the in-memory fakedriver): un-awaited delete reaches the driver with the collector off and on, and evicts a buffered set. Proven to fail against a lazy delete.
• test/pg-json-roundtrip.test.ts — pg-only (skipIf(!DATABASE_URL), own throwaway schema per test): snowflake string stays a string & get() agrees with the scan, nested JSON-looking payloads survive batchSet, and set(null)/batchSet(null) both store a jsonb null with exists() still true.

Verification

• pnpm check and pnpm typecheck clean; pnpm docs at zero warnings.
• Local run: 8 test files pass (57 tests) including the 3 new regression files, plus the pg-backed round-trip suite against a real DATABASE_URL.
• The two subprocess/SQLite exit-flush tests (exit-flush, idle-exit) fail only in this sandbox from a known better-sqlite3 native-ABI segfault; they exercise untouched machinery and are authoritative on CI (Node 22/24 ubuntu).

Tier 2 (NEW-1 / NEW-5 cross-process safety) and Tier 3 are deliberately out of scope for this release.

🤖 Generated with Claude Code