Where the disagreement starts

The claim that the fourth approach is merely a hybrid of methods 2 and 4, carrying every drawback of the latter, does not match the schema actually proposed. The intent is different, and the DDL makes it visible: shared vehicle data lives in a parent table, while car- and bicycle-specific columns live alongside a primary key that also references the parent.

All other tables are conventional: customer has an identity primary key, name and birthday are NOT NULL; vehicle carries manufacturer and price.

The subtype tables

car takes car_id as its primary key with a foreign key to vehicle, plus watt, doors, seats, powersource and license_plate. bicycle follows the same shape with bicycle_id referencing vehicle and a single gears column.

The contract

A contract points at a customer unconditionally (customer_id is NOT NULL) and at a vehicle through a nullable vehicle_id, with d_start, d_end and payment filling out the row.

Triggers, isolation and race conditions

The objection that triggers can never be a safe constraint mechanism because race conditions are unavoidable does not hold as long as they are written correctly. The difficulty described in that argument is not a trigger problem at all. It is the broader difficulty of expressing reliable constraints when running on lower isolation levels.

The follow-on concession stands: further variations and combinations of these techniques certainly exist beyond the ones listed, and the material itself is far from novel. Searching for "polymorphic associations" surfaces every one of these methods and more.

Whether a given approach is the best one is a question the author already hedged with "arguably." The real limitation is elsewhere: no method is perfect, and that imperfection traces back to how poorly some RDBMSes handle non-trivial declarative constraints.