Fully qualify T and NilClass references in generated RBI output

[sambostock]

Summary

Fully qualifies T references as ::T and NilClass as ::NilClass in all generated RBI output.

When a class defines a type member named T (e.g., via #: [T] in RBS), unqualified T::Boolean, T.nilable(...), etc. in generated RBI resolve to the type member instead of Sorbet's ::T module, causing type-checking errors. These types come from RBS translations (bool → T::Boolean, untyped → T.untyped, nil → NilClass) and unambiguously always mean the top-level constants.

Navigating the diff

Tip

Most of this diff is mechanical T → ::T and NilClass → ::NilClass string literal replacements. The interesting logic changes are in a handful of files:

Core qualification logic:

• lib/tapioca/helpers/rbi_helper.rb — Adds catch-all regex to sanitize_signature_types that qualifies T:: → ::T::, T. → ::T., and NilClass → ::NilClass. Updates as_nilable_type to emit ::T.nilable(...).

Places that called .to_s on Sorbet runtime types (producing unqualified strings):

• lib/tapioca/dsl/helpers/active_model_type_helper.rb — Applies qualification regex to type.to_s output
• lib/tapioca/dsl/helpers/graphql_type_helper.rb — Includes RBIHelper, routes return_type.to_s through sanitize_signature_types
• lib/tapioca/gem/listeners/sorbet_required_ancestors.rb — Routes ancestor.to_s through sanitize_signature_types
• lib/tapioca/sorbet_ext/generic_name_patch.rb — Routes type variable bounds .to_s through sanitize_signature_types

Avoided double-qualification:

• lib/tapioca/dsl/compilers/identity_cache.rb — COLLECTION_TYPE lambda updated to not blindly prepend :: when the type string is already qualified

Everything else in lib/ and spec/ is mechanical s/T::/::T::/ and s/T\./::T./ in string literals and test expectations.

Test plan

• All 784 existing tests pass (0 failures, 0 errors)
• Run tapioca dsl on a project with #: [T] generics to confirm the fix
• Verify Sorbet accepts the fully-qualified output

🤖 Generated with Claude Code