A guard reports its own failure and lets the call through

Type: bug — Track N. Found: 2026-08-18, jointly (frankonpiler-bf spotted the runtime throw; frank2-7e measured the boundary) while closing [[bug-n-from-sys-import-fails-while-import-sys-works]].
Measured at: HEAD 7e7ed35d7 and pinned v349 — same on both.

Repro

import sys
print(sys.version_info)

Unhandled exception: Exception: this build has no sys.version_info: the import it
came from could not be resolved, and the code guarding that (the flag its
except-branch sets) let this call through anyway

Two defects, and the second is the interesting one

sys.version_info is not provided. Other members are — sys.argv works, os.getcwd() works — so this is one absent member, not an unimplemented module. Providing it is a small decision with a question attached: what version does a NilPy build claim to be? Real code branches on it (html5lib/_tokenizer.py:21 is if version_info >= (3, 7)), so answering too low silently selects legacy paths. That part may want Track U.
The guard does not guard. The message states plainly that a flag set in an except-branch was supposed to stop this call and did not. A guard that detects its own failure, says so, and proceeds anyway is worse than no guard: it converts a compile-time refusal into a runtime crash in the user's program, and it has clearly been failing long enough for someone to write the sentence describing it.

Defect 2 is independent of defect 1 — the guard would still be broken for the next unprovided member. Fix it first; it is the one that generalises.

Resolved 2026-08-27 — defect 2. Defect 1 split to Track U.

The guard was not failing. There was no guard.

The ticket read the message literally and concluded a flag set in an except-branch was being ignored. Reading the code says otherwise: sys.<unknown> in pyparser.inc (~11229) deliberately emits a raise — "an attribute this sys does not have — raise, as CPython does" — and it is right to. It just emitted it through pyoptional_missing, the raise built for a different concept: a name bound to None because its optional import could not be resolved.

So one raise served two concepts, and its sentence described only the first. For import sys, which resolves perfectly and which no except-branch guards, every clause of that sentence is false. The message was not reporting a failure; it was borrowed. That is worse than the ticket's reading, not better: a sentence describing a mechanism that is not there is what sent this ticket looking for a broken flag.

The half the message hid

The wording was the visible defect. The load-bearing one is the class: pyoptional_missing raises a bare Exception, so

try:
    sys.version_info
except AttributeError:
    ...

— the idiomatic probe for an optional member, and the shape real code uses — walks straight past it and the program dies. Measured at pinned v379: the witness test does not print a caught message, it aborts on line 1. CPython raises AttributeError there, so this was a program CPython accepts and runs that NilPy could not, which is the exact definition of an N bug.

The fix

pylib.pas — pyattr_missing(owner, attr), raising AttributeError.Create('module ''sys'' has no attribute ''version_info'''). CPython's class, and CPython's sentence character for character.
pyparser.inc — PyMakeAttrMissingCall, and the sys.<unknown> site routed to it. Deliberately does not set PyExprHadOptionalMissing and does not call PySkipOptionalMissingChain: both belong to the optional-import concept. sys is a real module, a name assigned from it is a real variant, and .major on the result must compile to an ordinary runtime attribute access — CPython raises at the first step here too, so there is nothing to swallow.

pyoptional_missing keeps its message, which is true of the callers it still has (pyparser.inc ~14812/15230/15241, the unresolved-import family).

Defect 1 — `sys.version_info` is still absent, deliberately

Split to [[decide-nilpy-what-version-does-sys-version-info-claim]] (Track U, prio 62), exactly as this ticket predicted it would need to be. The implementation is fifteen minutes; the number is a product claim, because real code branches on it (if sys.version_info >= (3, 7)) and any answer silently steers third-party libraries down one path or the other. Nothing in the tree claims a Python version today, so this sets the precedent. Escalate, don't guess.

What ships today is a defensible resting place rather than a gap: the raise is catchable and correctly worded, so a program that probes copes and a program that reads it unguarded dies with CPython's own message.

Found while testing, NOT fixed here

getattr(sys, name) does not compile — error: undefined variable (sys), i.e. sys as a value. That is [[bug-n-a-resolved-module-member-as-a-value-is-an-undefined-variable]], already filed at prio 70; the witness test spells its probe without getattr and says why in a comment.

Gate

make compiler/pascal26 (fixedpoint 52e7b3b95cb5), tools/gate.sh quick GREEN, and a witness row test_nilpy_missing_module_attr_is_attributeerror in test-core whose .expected is CPython's own output. At pinned v379 it does not merely differ — it aborts on the first row, which is the whole point.

compiler/builtin/pylib.pas changed, so this needs a PIN.

Log

2026-08-27 — resolved, commit 0ecd624e6.