@dataclass: a str field's DEFAULT is silently dropped (becomes '')

from dataclasses import dataclass

@dataclass
class D:
    a: str
    s: str = "t"
    i: int = 7
    f: float = 1.5
    b: bool = True

d = D("x")
print(repr(d.s), d.i, d.f, d.b)

pxx:     '' 7 1.5 True
CPython: 't' 7 1.5 True

The int, float and bool defaults sitting right beside it are all correct, which is exactly what makes this hide: the class looks like defaults work.

Silent — no diagnostic, and '' is a perfectly ordinary value, so it surfaces far away as an empty name/path/label.

Confirmed pre-existing

Built from stable_linux_amd64/default/pinned as well as HEAD: both print ''. Not a regression from the repr work landed alongside this ticket (the control removes the variable — a pinned BINARY, not a source revert).

Where it is NOT — four candidates eliminated by measurement

This took four rebuilds to narrow, so the dead ends are recorded rather than left for the next session to re-walk. PXXDBG=n.ctorargs prints every construction's argument kinds and types and is the probe that answers this.

The call site builds the default as [1]kind=2,tk=4 — an AN_STR_LIT tagged tyString (frozen inline) — while the parameter it feeds is tyAnsiString (managed). That mismatch is the fault. The question was who builds that node.

The boundary — it is NOT "annotated str param with a default"

Varying the shape, every one of these is correct today:

def g(a: int, s: str = "t") -> str: ...      # correct
def f(a, s="t"): ...                          # correct
class K:
    def __init__(self, a: int, s: str = "t"): ...   # correct

So DefaultArgValueNode (parser.inc:2649) builds a tyString literal for an annotated str parameter in the ordinary cases and the conversion to a managed AnsiString parameter works there. Only the generated dataclass ctor's call site loses it. That is the remaining difference and where the next session should start: what PyClassCreate does to a default argument that a plain ParseCall does not.

Why it was not fixed in the same session

The remaining suspects are parser.inc and the argument lowering in ir*.inc — shared core files under the sole-A guard, and this session could not confirm sole-A. Same block as the four tickets behind decide-sole-a-guard-for-unattended-sessions.

Gate when it lands

Extend test/test_nilpy_dataclass_repr.npy, which deliberately uses no string defaults today and says why — pinning the wrong answer there would have frozen this bug. Add a str default to its Mixed class and regenerate .expected from CPython.

2026-08-09 — FIXED, and the earlier recon's four dead ends were all correct

PyClassCreate had a loop that filled every still-unsupplied STRING parameter with a HARDCODED EMPTY LITERAL (SOffset := 0; SLen := 0), and it ran before the dataclass-default loop. So it consumed the slot, and the default was never consulted. Moving it to run LAST — after the dataclass defaults, filling only what remains — is the whole fix.

That is why the earlier session could prove three separate default-builders were NOT in the path and still not find it. PyDcDefaultNode, PyClsAttrNode and DefaultArgValueNode were each markered with a distinctive AN_INT_LIT and each failed to fire — all three answers were correct, because by then the argument already existed. The eliminations were sound; the missing question was "who filled this slot BEFORE any of them ran?"

The generalisable version, worth more than the fix: when every candidate producer is proved absent, stop looking for a producer and ask what CONSUMED the slot. The n.ctorargs probe showed an argument present with a plausible shape, which reads as "someone built it wrongly" and not as "someone built it first".

Also confirmed by the same probe: the explicit and defaulted arguments are both AN_STR_LIT with tk=4, so the node's type tag was never the difference — an earlier hypothesis that cost a build to eliminate.

Verified against CPython: defaults given and omitted, several string defaults in one class, a string default before and after other kinds, an EMPTY string default (which must stay empty), and a hand-written __init__ with a string default as the control. test/test_nilpy_dataclass_repr.npy was also extended with the string default its own header promised to add once this was fixed — placed LAST in the class, since trailing string slots are what the bug consumed. gate.sh quick GREEN.