An annotation is a declaration here, a hint in CPython

Measured, one program per row

Why this is chosen and not a defect

An annotation is the ONLY way a NilPy field or local acquires a static type, and the compiler says so in its own voice — cannot infer the type of field self.a - annotate it (self.a: int = ...) is the diagnostic that half this lane's tickets are about. If an annotation did not determine storage, that advice would be advice to write something inert. The rule is also uniform: locals behave exactly as fields do, which is evidence of a design rather than of a field-inference accident.

Both answers are correct about their own implementation. CPython's annotations are metadata by specification; ours are a declaration in a dialect that compiles to machine code, and PXX's own architecture note — "compatible with FPC means the VALUE, not the intermediate's type" — is the same shape of call one layer down.

The residual question, and who owns it

self.a: int = t with t = 2.7 stores 2 with no diagnostic. That is a wrong value rather than a wider one, it is silent, and "silent" is what separates it from the other three rows. It follows from the same chosen rule, so it is recorded here rather than filed as a bug — but the reasoning that makes the float row obviously fine does not obviously cover it.

What would settle it: real source that annotates int and stores a float, correct under CPython and wrong under pxx. Absent that, this stays as it is; CLAUDE.md's own test — prefer the answer that leaves the mistake visible — arguably points at a diagnostic here rather than at CPython's answer. A narrowing conversion warning would satisfy both readings and change no value.

Found while probing every rung of PyInferFieldDecl for refactor-n-the-field-type-pre-pass-asks-one-question-in-six-places; not looked for.

A NARROWING WARNING IS COMPATIBLE WITH THE CHOSEN SEMANTICS (frankuser, 2026-09-09, and it sharpens the residual above rather than reopening it). "Chosen" covers the RULE — an annotation declares storage — and it does not cover the SILENCE. self.a: int = t with t = 2.7 loses a value with no diagnostic, and a warning at the narrowing store would satisfy both readings while changing no value and no type. Same shape as the closed-world case one lane over: a WARNING is compatible with the behaviour we want, a REFUSAL is not. Still not worth a fix without real source that wants it — a probe cannot settle whether anyone writes this — but if someone touches this area, the warning is the cheap half.