db-constraints (M5)

Constraints are invariants the database guarantees no matter which app, script, or migration writes the row. Without them, "this is always true" is a hope, not a fact. This module is design-axis (Constraints category). It applies to engines that enforce declarative constraints; document stores route to the Validación-schema category instead.

What it checks

• Missing NOT NULL: columns that are semantically required (FKs to mandatory parents, email, status, timestamps) left nullable, allowing partial/garbage rows.
• Missing CHECK: domain rules not enforced — quantity >= 0, price >= 0, status IN (...), start_date <= end_date, email ~ '@'. The app "validates" it; the DB does not.
• Missing UNIQUE: natural keys (email, username, slug, external_id) with no UNIQUE constraint, permitting duplicates that corrupt joins and auth.
• Case-insensitive uniqueness (emails): a plain UNIQUE on a case-varying column (e.g. email) still admits Foo@x.com and foo@x.com as distinct rows — duplicate identities. Name the remedy: a citext column type, or a UNIQUE expression index on lower(col) (e.g. CREATE UNIQUE INDEX ON users (lower(email))), so case-folded values collide.
• Over-nullable UNIQUE trap: a UNIQUE constraint over nullable columns where the engine treats each NULL as distinct, so duplicates slip through (e.g. UNIQUE(user_id, deleted_at) for soft-delete uniqueness fails when deleted_at is NULL). Recommend a partial/filtered unique index or NULLS NOT DISTINCT.
• Constraint declared NOT VALID / NOCHECK and never validated.

Axis & severity

• Axis: design; magnitude banded, never a fabricated violation count.
• Missing UNIQUE on an auth/natural key allowing duplicate identities: severity 4, fail/warn, confidence directional.
• Over-nullable UNIQUE that silently permits dupes: severity 4, warn (subtle, high-impact).
• Missing NOT NULL on a required FK/column: severity 3, warn.
• Missing CHECK for a stated domain rule: severity 2–3, warn, fixable: proposed.
• M5 does not hold a sev-5 cap; it shapes the Constraints category value.

Tier-0 static check

Parse DDL/snapshot via scripts/parse-schema.mjs: list columns lacking NOT NULL; detect natural-key column names without a UNIQUE/unique index; find UNIQUE constraints whose member columns are nullable (the over-nullable trap); scan for absent CHECK on quantity/price/date-range columns. Program-source parses stay directional.

Tier-1 verification query

Inventory constraints and nullability:

-- $DATABASE_URL, read-only
SELECT c.table_name, c.column_name, c.is_nullable,
       tc.constraint_type
FROM information_schema.columns c
LEFT JOIN information_schema.constraint_column_usage ccu
  ON ccu.table_name = c.table_name AND ccu.column_name = c.column_name
LEFT JOIN information_schema.table_constraints tc
  ON tc.constraint_name = ccu.constraint_name AND tc.constraint_type IN ('UNIQUE','CHECK')
WHERE c.table_schema NOT IN ('pg_catalog','information_schema');

Confirm the over-nullable UNIQUE actually has duplicates:

SELECT user_id, count(*) FROM memberships WHERE deleted_at IS NULL
GROUP BY user_id HAVING count(*) > 1;

Findings

Emit per schema/finding.schema.json. Examples:

• M5.users.email_no_unique — email has no UNIQUE constraint; duplicate identities possible (severity 4, warn, axis design, fixable: proposed).
• M5.memberships.over_nullable_unique — UNIQUE(user_id, deleted_at) lets NULL deleted_at duplicate (severity 4, warn, axis design).
• M5.order_items.quantity_no_check — no CHECK (quantity >= 0) (severity 3, warn).
• M5.events.user_id_nullable — required FK column is nullable (severity 3, warn). Each finding: evidence.observed quotes the column/constraint DDL verbatim; verification.reproduce is a runnable query above (method: ddl_parse / constraint_check / query_stat); expected_impact banded + confidence-tagged.

Honesty

• App-layer validation reduces but does not replace a DB constraint — multiple writers (jobs, admin scripts, replicas) bypass the app. State this plainly; do not call app validation equivalent.
• The over-nullable UNIQUE trap is engine-specific (Postgres pre-15 treats NULLs distinct; NULLS NOT DISTINCT since 15). Scope advice to the detected engine/version, do not generalise.
• Never claim duplicates/violations exist without a Tier-1 count; static, it is directional.