fix: reject non-Ref foreign key cycles at metamodel generation time

[zantvoort]

Fixes #412.

Problem

Two entities referencing each other with non-Ref @FK fields — or an entity referencing itself — compiled with no diagnostic, but the generated metamodels could never be used. Their constructors build record-typed children eagerly, so the metamodels of a cycle construct each other until the stack overflows: the first metamodel use after a clean build died with an ExceptionInInitializerError wrapping a StackOverflowError. The existing guards in the processors only prevented generating a file twice; the generated code itself still recursed. Verified with a reproduction test against the unmodified processor before applying the fix: the mutual Owner/Pet pair compiled cleanly and constructing the generated metamodel overflowed the stack.

Fix

Both processors (APT and KSP) run a memoized depth-first check over non-Ref record fields before generation and reject a cycle with an error naming its members and the fix:

Cycle of non-Ref foreign keys: Owner -> Pet -> Owner. A foreign key cycle must cross
a Ref boundary to be loadable. Mark one of the foreign keys as Ref (for example
Ref<Owner>) to break the cycle.

This mirrors the rule the engine already states at template level for self-references (TemplatePreparation): a foreign key cycle must cross a Ref boundary to be loadable. A cycle of inline records gets its own diagnostic, since an inline record embeds its columns and a cycle cannot be modeled at all. The error is attached to the field that closes the cycle and each cycle is reported once, not doubled by the nullable-chain variants.

The runtime record validation (RecordValidation.validateRecordGraph), which catches such a cycle at first model use when the processors are not attached, reports the same message: it renders the cycle members from the type the cycle re-enters and names the Ref fix, instead of the bare "Cyclic dependency detected". Its traversal already skipped Ref edges, so enforcement is unchanged; only the message is aligned.

Tests

Mirrored across both processor suites:

• the mutual non-Ref @FK cycle is rejected, with the cycle named and (Java side) the diagnostic asserted to print exactly once,
• a self-referencing non-Ref @FK is rejected (Employee -> Employee),
• a cycle that crosses a Ref boundary still generates as before,
• a three-entity chain closed by a Ref at depth generates as before, including the reference metamodel at the boundary and the navigation metamodels beyond it.

Full 25-module reactor build is green with the check active and produces zero cycle diagnostics, so no existing model is affected. The storm-core validation and metamodel test classes (155 tests) are green after the runtime message alignment.

[codecov]

Codecov Report

❌ Patch coverage is 50.00000% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...va/st/orm/core/template/impl/RecordValidation.java	50.00%	2 Missing and 3 partials ⚠️

📢 Thoughts on this report? Let us know!