delattr, globals, locals

Measured 2026-08-15 by sweeping the builtin surface one name per program, the same pass that found [[bug-nilpy-setattr-is-absent]] (fixed). All three fail LOUDLY at compile time as undefined variable, so nothing here is a silent wrong answer.

delattr(o, name) — a real gap

The completion of the attribute trio's write side. setattr now writes through pydynattr_set/_v; there is no pydynattr_del to call, so this needs a runtime entry as well as a frontend arm. Note the neighbouring refusal in pyvar_delitem, which declines del v[k] on a user object for the same reason — a dispatcher that does not exist yet — and says so loudly rather than silently doing nothing. Same trade applies.

Worth doing together with del o.attr, if that turns out to be absent too: measure it before assuming.

globals() / locals() — probably a divergence, not a bug

Both want a run-time name table mapping source names to storage. NilPy compiles locals to stack slots, and exec(src) with no namespace is ALREADY refused by name for exactly this reason, with a message that explains it and points at the explicit-dict form. globals() is the more tractable half (module-level names do have storage), but a dict that does not write through is worse than no dict at all — a globals()["x"] = 1 that silently does nothing is the failure mode this project refuses.

Escalate rather than guess if either is wanted: it is a devdocs/dev/nilpy-semantics-divergences.md entry or a Track U decide-*, not a quiet implementation choice.

Gate

Per name as it lands: a .npy diffed against CPython, plus a row asserting a user def delattr still shadows the builtin.