refactor(db): migrate ORM from Drizzle to Prisma

[jee7s]

Summary

Replaces Drizzle ORM with Prisma 7 across @arbor/db, preserving the public store contract so no consuming package changes: PostgresUrlStore / PostgresConfigStore / PostgresAuditStore (and SQLite equivalents), the createStores() factory, and the UrlEntry / AuditRecord types are all unchanged. The agent, lambda, logger, and mcp-url-fetcher packages are untouched.

Dual-database design (preserved)

The existing "Postgres in prod, SQLite for dev/tests" split is kept. Prisma binds a client to one provider at generate time, so this uses two schemas generating two clients:

	Schema	Generated client	Driver adapter
Postgres (prod)	prisma/postgres/schema.prisma	generated/postgres	@prisma/adapter-pg
SQLite (dev/test)	prisma/sqlite/schema.prisma	generated/sqlite	@prisma/adapter-better-sqlite3

createStores() selects the client by connection string exactly as before. Prisma 7 has no embedded query engine (it uses a WASM query compiler + driver adapters).

SQLite specifics

The zero-migration, schema-on-first-connect behaviour is retained — including :memory:. The better-sqlite3 adapter keeps a single connection, so the bootstrap DDL (run via $executeRawUnsafe) applies to the same connection the client queries. timestampFormat: "iso8601" preserves the ISO-string added_at / created_at contract.

Tooling changes

• db:generate → prisma generate for both schemas. A postinstall hook regenerates the clients on every npm install/npm ci, so CI's non-build jobs (unit, dev) and Docker always have them. Generated clients are gitignored (packages/db/generated/).
• db:migrate → prisma migrate deploy. Postgres migrations now live in prisma/postgres/migrations/; the baseline 20260608000000_init mirrors the previous Drizzle schema.
• Dockerfile: migrate stage runs prisma migrate deploy; the builder copies schemas before install (for the postinstall generate) and the runtime stage copies the generated clients.
• Removed deps: drizzle-orm, drizzle-kit, postgres. Added: @prisma/client, prisma, @prisma/adapter-pg, @prisma/adapter-better-sqlite3.

⚠️ Operational note — baseline existing databases once

Existing prod/dev databases already have these tables. Before the first migrate deploy, mark the baseline as applied so Prisma doesn't try to recreate them:

cd packages/db
DATABASE_URL="postgres://..." npx prisma migrate resolve \
  --applied 20260608000000_init --config prisma.config.postgres.ts

Fresh databases need no baselining. Documented in docs/deployment.md.

Verification

• ✅ Full monorepo build (npm run build)
• ✅ 255 unit tests (npm run test)
• ✅ 27 SQLite dev/integration tests (npm run test:dev) — :memory:, file-backed persistence, Boolean/DateTime/ISO roundtrip, autoincrement ids, ordering, null model
• ✅ Coverage thresholds still met (91% lines / 87% branches)
• ✅ Lambda bundle: WASM query compiler inlined as base64 (no on-disk .wasm), bundled handler loads, Postgres query path initialises (connection-only failure against a dead DB, i.e. the compiler works)
• ⚠️ Live Postgres roundtrip not run locally (no Postgres server in the dev env); the query layer is identical to the SQLite path that passes, and CI's Postgres integration job exercises migrate deploy against a real DB

Notes for reviewers

• Prisma model fields are camelCase mapped to the existing snake_case columns (@map), so the DB schema is unchanged.
• delete uses deleteMany to keep the no-throw-on-missing semantics Drizzle had.
• This branch is independent of the ECR→GHCR branch; the only workflow changes here are one-line comment fixes (drizzle-kit migrate → prisma migrate deploy).

🤖 Generated with Claude Code