Release v2.0.3

Summary

Interpreting a capture without a language model (cratis prologue interpret --no-llm) now produces a model that passes executable validation, not only parsing.

Fixed

• Each inferred command produces only the events its own correlated transaction recorded, also when similar routes such as nested /owners/{id}/pets/new are merged into one slice.
• An event used by several slices is declared once, so models no longer report it as ambiguous or as a duplicate stored name.
• An insert followed by an update of the same table inside one transaction is read as one creation instead of a separate edit.
• Required numeric properties, such as foreign keys, no longer get a text-only not empty rule.
• Every read model that a generated query or projection uses is declared, keyed explicitly on its row identity. (#46)