raise KeyError(42) segfaults

try:
    raise KeyError(42)
except KeyError as e:
    print(e.args)

Segfaults at the raise. PRE-EXISTING — identical under stable_linux_amd64/default/pinned.

• Found: 2026-08-13, sweeping e.args for [[bug-nilpy-exception-args-attribute-missing]] with a non-string key.
• Silent and fatal, which is the worst pair: no diagnostic, no exception, just a dead process on a program CPython runs.

Cause, as far as measured

Exception.Create takes const m: AnsiString and every builtin exception inherits it. A NilPy raise Cls(x) passes the argument straight through, so an integer arrives where a managed string handle is expected and the first read of its length word walks off a small integer. That is the same shape as every other "a scalar reached a string parameter" crash in this frontend.

Note the multi-argument form does NOT crash: raise MyErr("a", 42) is folded to a rendered string at the construction site, and the fold boxes each argument through pyvar_repr. So it is specifically the ONE-argument non-string raise that has no conversion.

Shape of a fix

At the raise/construction site, box a non-string single argument the way the multi-arg fold already does — pyvar_repr for the message — and keep the raw value for args (the argsv override that landed with the args ticket is exactly the slot for it). That makes str(KeyError(42)) be 42 and e.args be (42,), both matching CPython.

Alternatively refuse it at compile time, which is strictly better than the segfault but loses an ordinary Python idiom (an integer error code as the payload).

Gate

A .npy diffed against CPython: raise KeyError(42), a float and a tuple argument, str(e), repr(e) and e.args for each, plus the string forms as controls.

Log

• 2026-08-13 — resolved, commit beacc6660.

Resolution (2026-08-13)

Fixed as the ticket's first option — box at the construction site — in two shapes, because the family is not uniform:

• KeyError takes a Variant message now and reprs it itself. That makes str(KeyError(42)) be 42 (unquoted, CPython's rule for the one builtin whose str() is repr(arg)) and e.args be (42,) with the key's own TYPE. KeyError() with no argument fills None and renders as ''.
• every other exception keeps Create(const m: AnsiString); the frontend boxes the single non-string argument and renders it through the new pyexc_msgstr (= str(), ONE signature so FindProc cannot pick a wrong overload — that is how the multi-arg fold crashed while fixing a crash).

Both fire only at nArgs = 1, so they cannot overlap with the multi-argument fold above them, which is untouched.

Residual, measured, deliberately shipped: e.args for a NON-KeyError exception is ('42',) where CPython says (42,) — that family's message parameter is still a string, and widening the RTL's base Exception.Create to a Variant would change every raise EFoo.Create('..') in the RTL under a quick-only gate. Filed as [[bug-nilpy-non-keyerror-exception-args-loses-the-argument-type]]. str() and repr() are exact for every shape tested (int, float, bool, tuple, string, empty, user subclass, multi-arg, and a dict miss).

Verified against CPython by test/test_nilpy_exception_non_string_argument (wired into test-nilpy) plus the six existing exception tests re-run by name.