repr() and type(e).__name__ leak the internal PyException name

Measured at HEAD:

Only the ROOT is affected — every named subclass declares its own Python spelling, so ValueError, KeyError and the rest are correct. That is exactly why the existing tests missed it: test_nilpy_exception_args asserts ValueError('v') and KeyError('inner'), and nothing asserted a bare Exception.

Cause, and it is self-inflicted

[[decide-pylib-exception-vs-sysutils-exception]] option 5 renamed pylib's root class to PyException so it would stop colliding with sysutils'. ClassName (and therefore repr and type(e).__name__) reports the declared name of the class, and the declared name is now the internal one. The lexer maps the bare identifier on the way IN; nothing maps it back on the way OUT.

Why it matters

if type(e).__name__ == "Exception": and error text that embeds repr(e) are ordinary Python. Under NilPy's rule — if code works on CPython, it must work on NilPy — this is a real break, not a divergence to document.

Fix

Fixed by construction by [[feature-a-one-exception-class-in-a-shared-unit]]: the shared class is NAMED Exception and pylib reaches it through an alias, so ClassName is Exception again. Verified in the prototype for that ticket (bare root ClassName: Exception).

If that lands, close this with it. The standalone alternative — special-casing the name in the renderer and in __name__ — is a second place that has to know the mapping, and mapping-in-one-direction-only is what caused this.

Gate

repr(Exception('x')), str(Exception('x')), type(e).__name__, and a subclass's __name__, all diffed against CPython. Add the bare-root rows to test_nilpy_exception_args.npy — their absence is why this shipped.

2026-08-14 — ALREADY FIXED on a branch. Do not start this from scratch.

feature-a-one-exception-class-in-a-shared-unit (variant C) fixes this by construction rather than by another rename: pylib's root is named Exception again, as a sibling of sysutils' class of the same name under a shared ExceptionBase, so ClassName reports Exception because that IS the declared name. Measured green on wip/exception-sibling-design for ValueError, KeyError, a user class MyErr(Exception), and ValueError(42).args[0] + 1.

That branch also DELETES the pylexer Exception -> PyException mapping, whose "maps in, never maps out" asymmetry is what produced this bug in the first place — so the fix removes the mechanism rather than adding a second rename to compensate for the first.

Blocked on [[decide-merge-variant-c-with-bare-name-collision]], not on effort. Picking this ticket up independently would either duplicate that work or add the compensating rename the branch exists to delete.

Fixed 2026-08-14 by merging variant C — exactly as this ticket prescribed

Landed with [[feature-a-one-exception-class-in-a-shared-unit]] (squashed from wip/exception-sibling-design), per the user's option-A decision on [[decide-merge-variant-c-with-bare-name-collision]]. Fixed by construction, not by a second rename: pylib's root is named Exception again, so ClassName — and therefore repr() and type(e).__name__ — reports the right name because that IS the declared name.

The pylexer Exception -> PyException mapping is deleted. Its "maps in on the way, never maps out" asymmetry is what produced this bug, so the fix removes the mechanism rather than adding a compensating one. That was this ticket's own argument against the standalone alternative, and it held.

Measured against the CPython oracle — identical on every row

repr(Exception('x'))   Exception('x')        (was PyException('x'))
str(Exception('x'))    x
type(e).__name__       Exception             (was PyException)
repr(ValueError('v'))  ValueError('v')       unchanged
class MyErr(Exception) MyErr                 unchanged
except Exception as e  Exception             binds and names correctly
ValueError(42).args[0] + 1  ->  43

The Gate's test rows — added, and controlled

test_nilpy_exception_args.npy now carries the bare root: repr, str, __name__, args, a bare raise/catch, and a subclass's __name__ (which must still be the subclass). Expectations regenerated from CPython.

Their absence really was why this shipped, and that is now demonstrated rather than asserted: run against the PINNED pre-merge compiler, the file differs on exactly the two new lines —

-bare root Exception('x') x Exception ('x',)
-bare cat  Exception Exception('raised') ('raised',)
+bare root PyException('x') x PyException ('x',)
+bare cat  PyException PyException('raised') ('raised',)

— and nowhere else, because every pre-existing row uses a named subclass that declares its own Python spelling.

Two things the branch missed, found by grepping for the old name

Reading the branch's diff would not have surfaced either; both were caught by searching for PyException across everything buildable.

1. lib/pcl/tkinter.pas declared TclError = class(PyException) and no longer compiled. gate.sh quick cannot see this — it is a lib/pcl unit built with $(PXX_STABLE) — so after the pin every Track B build would have hit it. The pinned compiler was the control: it builds tkinter, HEAD did not. Now class(Exception), unambiguous because that unit's uses names pylib and not sysutils.
2. The Makefile's expectations for the two uses-order tests were stale. The branch rewrote those tests to assert the QUALIFIED property; the recipes still asserted the old bare-name output, in both copies of the recipe.

test_nilpy_pyexception_bare_vs_qualified.npy passes unchanged — its assertions were always about the observable contract, and only its prose described the deleted mechanism. That prose is updated; the assertions are not, which is what makes it the regression test for swapping the mechanism underneath.

Whole exception family re-checked against CPython: empty_exception_subclass, except_as_binder_scope, exception_args, exception_fstring_message, exception_multi_arg, exception_no_leak — all match. rtl_exception_surface matches its recorded expectation (it imports sysutils, so CPython cannot be its oracle).

Gate

gate.sh quick GREEN + self-host fixedpoint; make lib-test GREEN against the new pin (~80 suites, including the tk-nilpy one that exercises tkinter); sysutils' own Exception surface and named descendants (EConvertError) re-checked.

Re-pinned to v301 as part of the landing, which was required rather than tidy: sysutils now uses exceptions, and the old pinned compiler resolved that unit only when cwd happened to be the repo root — from anywhere else it failed with unit source not found: exceptions. Verified cwd-independent on the new pin, and in a fresh git clone.

One trap for the next person who adds a builtin unit: make pin prints git add -u stable_linux_amd64/ … all stable files are tracked, so nothing can dangle. That holds only while the frozen builtin SET is unchanged — -u does not stage a NEW file, so the first pin after adding a builtin unit ships a stable tree missing a source. Caught here by listing the frozen directory against git status; fixed in a follow-up commit.

Log

• 2026-08-14 — resolved, commit fb2e88334.