Make impossible relationships unrepresentable

First, build a mental model

A schema is a collection of promises about every possible saved row. A foreign key connects identities; a composite foreign key can connect identity AND company.

1. A primary key identifies a row. A foreign key requires a referenced key to exist; it does not infer that two independent company columns agree.
2. NOT NULL rejects missing values. CHECK rejects false, while a NULL result does not violate a PostgreSQL CHECK; pair the two when missing values are invalid.
3. The ordered columns of a composite foreign key must match a unique or primary key in the parent. This encodes the relationship rather than trusting a browser filter.
4. The example runs in the SQLite lab with foreign keys enabled. PostgreSQL also supports these constraints; schema evolution and locking differ between engines.

CREATE TEMP TABLE depots (
  company_id TEXT NOT NULL, id INTEGER NOT NULL,
  PRIMARY KEY (company_id, id)
);
CREATE TEMP TABLE bins (
  company_id TEXT NOT NULL, depot_id INTEGER NOT NULL,
  units INTEGER NOT NULL CHECK (units >= 0),
  FOREIGN KEY (company_id, depot_id)
    REFERENCES depots(company_id, id)
);
INSERT INTO depots VALUES ('company-a', 7);
INSERT INTO bins VALUES ('company-a', 7, 3);

1. Parent

   The parent identity is the pair (company-a, 7), not merely the number 7.

2. Child

   The valid bin points at that exact pair and contains a non-null, nonnegative count.

3. Counterexample

   Changing only the bin company to company-b produces a missing parent pair, even though depot 7 exists elsewhere.