Upgrade ty to 0.0.59

[bradhilton]

Summary

• upgrade the repository-wide type checker from ty 0.0.14 to 0.0.59
• fix newly exposed annotation and narrowing defects where the runtime contract is expressible
• keep intentional suppressions narrow and portable across mypy/pyright/ty where possible
• keep local type checking usable without optional Megatron dependencies

Diagnostic inventory and treatment options

The CI-shaped Megatron environment initially reported 276 diagnostics. The largest rule buckets were 139 invalid-argument-type, 28 unresolved-attribute, 26 invalid-assignment, 18 missing-typed-dict-key, and 12 unknown-argument.

1. Legacy suppression codes

   • Chosen: remove suppressions through better contracts and narrowing first; where an intentional escape hatch remains, use portable # type: ignore rather than paired mypy/ty comments.
   • Ty-specific directives remain only for six third-party bridge/DSL locations whose diagnostics are specific to ty's current stubs or static model.
   • Alternatives: retain paired checker-specific directives (more precise per checker, noisy), or disable affected rules globally (smallest diff, materially weakens checking).

2. Intentional test doubles and negative tests

   • Chosen: type structural fakes at their construction boundary, add runtime narrowing, and use real protocols for the tokenizer surface.
   • Alternatives: construct heavyweight runtime classes in unit tests, or add per-file rule overrides that hide unrelated future mistakes.

3. Real annotation/narrowing gaps

   • Chosen: fix them. This includes the structural backend contract, covariant eval results, optional packed-tensor fields, concrete accumulators, tokenizer capabilities, dynamic framework patch boundaries, subprocess environment typing, and numeric return normalization.
   • Alternatives: cast at every use site (less code movement, weaker evidence), or suppress each report (behavior-preserving but leaves inaccurate contracts).

4. Dynamic framework-extension boundaries

   • Chosen: use getattr/setattr with runtime validation for deliberately installed methods, modules, tensor metadata, and test mutations.
   • Alternatives: local facade layers (cleaner but substantial), or broad module overrides.

5. Triton/TileLang and Megatron Bridge APIs

   • Chosen: retain narrow ty-only directives for the two GDN kernel launch modules and four bridge registrations whose JIT/decorator interfaces are not represented correctly by static signatures.
   • Alternatives: maintain local/upstream stubs or a checker plugin (best long term), exclude kernel files entirely, or globally ignore argument diagnostics (both broader than warranted).

6. Optional dependency environments

   • Chosen: allow unresolved optional imports in lightweight environments, while separately validating against the Megatron-populated environment.
   • Alternatives: require all heavyweight extras for every local type check, or maintain separate ty configuration files per CI lane.

No global type-error rule is disabled.

Validation

• ty 0.0.59, lightweight backend/dev environment: all checks passed
• ty 0.0.59, CI-shaped Megatron dependency environment: all checks passed
• prek: ruff, ruff-format, ty, and lockfile hooks passed
• non-Megatron unit suite: 376 passed, 11 skipped
• focused changed-path suite: 176 passed
• lockfile and diff integrity checks passed

Three Megatron-importing unit modules cannot collect on this CPU workstation because Transformer Engine cannot load its packaged CUDA runtime; the Megatron CI lane remains the runtime authority for those modules.