A multi-argument raise loses its argument tuple to the fold

Measured 2026-08-14 at HEAD, after [[bug-nilpy-non-keyerror-exception-args-loses-the-argument-type]] fixed the single-argument case:

class MyErr(Exception): pass
try:
    raise MyErr("no such user", 404)
except MyErr as e:
    print(e.args, len(e.args), str(e))

So code = e.args[1] next to raise MyErr("no such user", 404) — the exact idiom args exists for — reads the wrong thing, and len(e.args) is 1 for every multi-argument exception in the language.

Cause

Exception takes ONE parameter, so the multi-argument arm in pyparser.inc folds the arguments into the single message CPython would print — '(' + repr(a) + ', ' + ... + ')' — at the construction site. args is then derived from (or stored as) that one rendered string.

The single-argument case was fixed by widening the base ctor to a Variant, and that does not extend here: the problem is not the argument's TYPE, it is that there is one parameter and N arguments.

Shape of a fix

The fold already visits every argument and boxes each through pyvar_repr. It needs to ALSO build a TPyList marked PYSEQ_TUPLE from the same arguments and stash it in the argsv slot the base class already has (KeyError uses it; the storage and GetArgs' preference for it are both in place).

The obstacle recorded when args first shipped: the fold consumes its argument nodes into the rendered string, so building the tuple as well needs clones of those nodes — the same argument evaluated twice would be wrong for anything with a side effect, so it wants a hoisted temp per argument, evaluated once, feeding both the repr and the tuple. That is the real work here.

Gate

The table above matching CPython, e.args[1] == 404, the single-argument and KeyError rows of test_nilpy_exception_args and test_nilpy_exception_non_string_argument unchanged, and a multi-argument raise whose arguments have side effects evaluated exactly once.

Resolution (2026-08-15)

Fixed by inverting the order: build the tuple at the construction site, then derive the message FROM the tuple at run time. That dissolves the obstacle the "Shape of a fix" section records — no clones and no per-argument temps are needed, because each argument is now parsed and lowered exactly once, into the tuple, and the rendered text is a pure function of the finished tuple.

• compiler/builtin/pylib.pas: pyexc_tuplemsg(t) returns pylist_repr(t) — the same ('no such user', 404) text the fold used to concatenate by hand — and pyexc_setargs(e, t) stores the tuple into the argsv slot the base class already had.
• compiler/pyparser.inc: new PyMakeTupleFromArgs(firstArg) builds a PYSEQ_TUPLE TPyList from an already-parsed AN_ARG chain via a hoisted temp (pylist_mark_tuple + append_self). The multi-argument fold now calls it, passes pyexc_tuplemsg(tup) as the ctor's one argument, and wraps the finished node in pyexc_setargs, keeping ASTRight so the record identity survives.

PyMakeTupleFromArgs is used ~6800 lines above its definition; pxx accepts that, FPC does not, and the seed canary caught it. Forward declaration added beside PyUserNameShadowsHere's — third instance of this shape this session, after bug-a-fpc-seed-drift-emitasmx64-forward.

Verified byte-identical to CPython, including the side-effect row that motivated the ticket:

args    ('no such user', 404) 2
index   no such user 404
istuple True
args3   ('a', 1, 2.5) 3
sides   ('x', 7) ['x', 7]      <- each argument evaluated exactly once
args1   ('solo',) 1
args0   () 0
argsbi  ('v', 2) 2
argskey ('nope',) 1 'nope'

test/test_nilpy_exception_multi_arg.npy extended in place (+.expected refreshed). All twelve pre-existing test_nilpy_(exception|raise|except)* tests re-diffed against CPython and unchanged.

Gate: make compiler/pascal26 fixedpoint + tools/gate.sh quick GREEN (self-host, testmgr quick, FPC seed canary).

Log

• 2026-08-15 — resolved, commit d135e88d1.