refactor(db): Prisma → Drizzle — the fleet has one ORM

[catomean]

Final repo of the stack-uniformity migration (George: "Drizzle everywhere") — after biaslens#26, reparaturbonus-zh#130 and solon#136. This was the largest of the four: 46 tables / 69 enums / 29 migrations at branch time, ~500 Prisma call sites in 116 src files, 62 test suites mocking the Prisma client — and, uniquely, a mid-flight master feature (#152) that had to be ported on the way through.

Recipe deltas vs reparaturbonus#130

• Schema by introspection, not by hand: scratch Postgres 16 built by the full Prisma migration chain, then drizzle-kit pull — exact camelCase columns, "AgeRange"-style enum names, TIMESTAMP(3), and every Prisma constraint/index name pinned by construction. Manual passes on top: mode: 'date', $defaultFn(createId) on exactly the 43 cuid ids, $defaultFn/$onUpdateFn on exactly the 29 @updatedAt tables (both sets diffed against schema.prisma), a bytea customType for the two Bytes columns, ARRAY[]::TEXT[] defaults restored (drizzle-kit misparses them as ["RAY"]), and per-column .op() specifiers stripped (drizzle-kit scrambles them positionally in multi-column indexes; all were Postgres defaults).
• Type-level FK cycles (Placement↔Incident, HouseRule↔Proposal) break with (table): PgTableExtraConfigValue[] => return annotations — AnyPgColumn casts alone don't cut the inference loop.
• Nullable TEXT[] truth: Prisma typed scalar lists string[]; the columns are nullable. The UI contract stays non-null via ?? [] at the projection boundary (toResidentUiSummary, ResidentSummary).
• Relations mirror Prisma's field names (photo, expensesPaid, taskRequestsReceived, …, relationName = Prisma's @relation names), so every include: became the same-shaped with:.
• types.ts reproduces what @prisma/client exported: row types AND per-enum runtime objects (StaffRole.ADMIN), derived from the pgEnums.
• isUniqueViolation() replaces PrismaClientKnownRequestError/P2002 (SQLSTATE 23505, unwraps DrizzleQueryError cause).
• site-access (Which places a staff member is responsible for #152) port: the where-fragment helpers now emit drizzle SQL; the residents-in-my-units rule (placements: { some: … }) is an inArray subquery built on a standalone QueryBuilder so constructing a filter never touches the lazy client (jest + next build run without DATABASE_URL); empty assignment renders false (drizzle inArray throws on []).

Parity — proven, not asserted

Scratch Postgres 16: full Prisma chain (30 migrations) → db A; drizzle/0000 + drizzle/0001 via drizzle-kit migrate → db B; normalized pg_dump --schema-only --no-owner --no-privileges diff (bookkeeping tables excluded): empty — 1486+ DDL lines, all tables, enums, indexes, FKs, uniques and the Message_one_author CHECK identical including names.

The live aoz_wohnen dump was also diffed against db A: semantically identical; only column/enum-value ordering differs (columns added by later ALTERs sit at the end on live; LOW_SATISFACTION was enum-appended) plus live's User_code_key being a unique index rather than a unique constraint — equivalent objects, append-history artifacts, no action needed.

Verification

• npm run verify green: format:check + lint + typecheck + 204/204 suites, 3584 passed / 8 skipped (3592). Baseline before migration: 202 suites / 3563+8; the deltas are master's two new suites (A coach should not be the last to know their client's house is in trouble #150–152) and dynamically discovered it.each rows (a source-scanning gate now finds 7 insert sites instead of 5) — no hand-written test was added, deleted, skipped or weakened.
• Hermetic npm run build green (no codegen step left; build is just next build).
• Seeds proven by execution: on a scratch DB, db:migrate + db:seed + db:seed:admin all exit 0 (25 residents, 18 placements, live-computed compatibility scores), and a second run over the populated DB confirms FK deletion order + idempotent governance path.
• CI rewired: no prisma generate steps; e2e schema init is npm run db:migrate, seeds are db:seed/db:seed:admin.

Live-DB cutover — non-destructive by construction, already armed

Deploys run fleetcrown's apply-schema.sh, which probes ./drizzle first. Both migrations are marked applied, never run, in both ledgers, on the box, before this merges — so the merge deploy's schema step logs "up to date" and ships code only:

• aoz_wohnen (live, 19 residents): _deploy_schema_history = {0000, 0001}; drizzle.__drizzle_migrations carries both sha256s + journal timestamps. The Which places a staff member is responsible for #152 Prisma migration had already reached this DB via master's deploy, so its schema is the 0000+0001 shape (verified above).
• aoz_demo (dormant demo instance — service disabled since 2026-08-26): was 9 migrations stale; the missing, verified-additive Prisma migrations were applied by hand first (15 residents before and after), then the same dual-ledger baseline. scripts/deploy-demo.sh now applies drizzle/*.sql through the same _deploy_schema_history ledger instead of npx prisma migrate deploy.
• After the first Drizzle deploy is verified live: DROP TABLE "_prisma_migrations" (step 3, run separately).

Row counts for every table were recorded before the cutover and will be re-verified after the deploy (nothing in this PR writes to the live DB; the deploy applies no schema).

Prisma is gone

prisma/ (schema + 30 migrations), @prisma/client, prisma, the package.json#prisma block, all CI generate steps, and every import. grep -rni prisma in src/scripts now hits only historical comments explaining why code is shaped the way it is. Guard tests that used to regex schema.prisma read the pgEnum objects at runtime instead. Docs (CLAUDE.md, README, INFRASTRUCTURE.md, demo guides) updated. Closes the door on dependabot #126/#127.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WqKqMnHQHSmkGFfc5t7Rxn