7. One module per procedure

Context

The catalogue is hundreds of small, independent procedures with no shared state and no ordering dependencies. That structure invites parallel implementation — several agents or contributors working on different procedures at once.

Decision

Exactly one procedure per module, and the module name must equal the procedure id. This is enforced at registration: register raises if they disagree. There are no collector modules — no __init__ importing procedures by name, no shared per-procedure test file. The three registry-wide test suites iterate whatever autodiscovery finds.

Consequences

Adding a procedure touches only files that are new, so parallel work never produces a merge conflict. The cost is that discovery is dynamic rather than explicit, so a typo in a module name shows up as a missing procedure rather than an import error — which the invariant suite catches instead.