fix(cpp): a nullptr argument picks the std::nullptr_t overload, and no call links to a deleted overload

[colbymchenry]

Summary

Two ways a C++ call reached an overload that no compiling call reaches:

• nullptr against std::nullptr_t. cppArgumentTypeFit scored a nullptr literal 3 against any pointer parameter. It scored only 1 against a std::nullptr_t parameter, whose type counted as a class. protocolbuffers/protobuf's Arena::UniquePtr<T> declares reset(), reset(std::nullptr_t) and reset(T* PROTOBUF_NULLABLE) = delete. So ptr.reset(nullptr) took the pointer overload. C++ takes reset(std::nullptr_t), an exact match, over a pointer conversion.
• Deleted overloads. That reset(T*) is deleted, and so is RepeatedPtrField<Element>::Add(const Element&), beside Add(Element&&). Eighteen field.Add("foo")-style calls in protobuf's tests reached the deleted one. When overload resolution in C++ selects a deleted function, the program is ill-formed, so no compiling call reaches one.

This came out of the interaction A/B for the #2536–#2542 batch, where reset(nullptr) was the one wrong composed result.

What changed

All in name-matcher.ts:

• cppTypeCategory: std::nullptr_t has a category of its own, null.
• CPP_LITERAL_FIT:
  • nullptr scores 4 against std::nullptr_t, its own type, above a pointer's 3;
  • a literal 0 scores 1 against it, below an integer's 2;
  • a character, number or bool can't become one (-3).
• cppOverloadFor: overloads declared deleted are dropped, unless every overload is. isCppDeletedFunction reads a declaration from its name to its ; or body for = delete, memoized per node and cleared with the other C++ memos.

Validation

Site-keyed edge diffs of main 24286218 against this branch. Both arms were built from one tree, kernel included.

repo	removed	added	retargeted	metadata-only
google/leveldb	0	0	0	0
fmtlib/fmt	0	0	0	0
protocolbuffers/protobuf	0	0	19	0
facebook/rocksdb	0	0	0	0
redis/redis (C, control)	0	0	0	0

Node and file counts are identical. All 19 are right:

• ptr.reset(nullptr) (arena_unittest.cc) reaches reset(std::nullptr_t) instead of the deleted reset(T*).
• 18 RepeatedPtrField::Add calls in repeated_ptr_field_unittest.cc reach Add(Element&&), by its out-of-line definition, instead of the deleted Add(const Element&). They include field.Add("foo"), field.Add(std::move(test_all_types)), rep.Add(v) on a const char*, and elements.Add(absl::StrCat(…)).

Tests

__tests__/cpp-nullptr-deleted-overloads.test.ts, 3 tests:

• protobuf's shape: ptr.reset(nullptr) reaches reset(std::nullptr_t) and ptr.reset() reaches reset(). With both fixes off, as on main, it fails.
• Set(nullptr) against Set(Widget*), Set(std::nullptr_t) and Set(int), with nothing deleted. It reaches Set(std::nullptr_t), and Set(0) reaches Set(int).
• Scale(2) against a deleted Scale(double), Scale(int), and a deleted Scale(long) declared over two lines. It reaches Scale(int).

Mutation check (each guard broken alone; every one fails a test, at its own assertion):

guard	fails
std::nullptr_t is its own category	Set(nullptr)
nullptr fits std::nullptr_t over a pointer (4 vs 3)	Set(nullptr)
a literal 0 fits an int over std::nullptr_t	Set(0)
a deleted overload is taken only when every one is	Scale(2)
a declaration is read only to its own ; or body	Scale(2) (Scale(int) would read the deleted Scale(long) after it)

In protobuf's case either fix alone is enough, so each has a fixture only it decides.

Full suite: 7,047 passed, 82 skipped. The 9 failures, in 3 files and all from load on this machine, pass when those files run alone (24/24).

🤖 Generated with Claude Code