fix(cpp): a -> call on a declared container iterator calls the element's method

[colbymchenry]

Summary

This is the follow-up left after #2573.

A -> call on a C++ receiver declared as a container's iterator was matched by the method's name alone. Examples: std::deque<Slice>::const_iterator it; it->data(), and std::vector<std::string>::const_iterator itr; itr->c_str(). Such a call could reach any class with a method of that name: in rocksdb, itr->c_str() reached gtest's FilePath::c_str.

A -> on an iterator now reaches the container's element, the same way -> on a smart pointer already reaches what it holds (matchCppHeldTypeCall):

• A new cppArrowTarget reads the element from the container's template arguments (cppRangeElementOf). It does this before the container's own iterator alias is followed, because following it would lose those arguments: protobuf's RepeatedPtrField<T>::iterator is a typedef of internal::RepeatedPtrIterator<Element>.
• So the -> branch now also takes declarations read through an alias. Holders are unaffected, since cppHeldType only matches the literal std::unique_ptr<…>-style spellings.
• The element is looked up as C++ does.
• A library element (std::string) or the caller's own template parameter calls no project method, as with std::unique_ptr<T>.
• Some elements are written only by a class template's own begin()/operator*, using its parameter (googletest's ParamGenerator<T>::iterator). Only the specialization decides those, so their calls go on as before.

Validation

• New test __tests__/cpp-iterator-arrow.test.ts (5 tests): 4 fail on main and pass here. The ParamGenerator guard passes on both.
• Mutations: switching off each part fails its test:
  • the iterator target: 4 tests;
  • taking aliased declarations: 1;
  • the class-template guard: 1;
  • the caller's-own-parameter branch: 1;
  • reading the container before the alias: 1.
• A/B (full index, main vs fix) on google/leveldb, fmtlib/fmt, redis (C control), facebook/rocksdb and protocolbuffers/protobuf. Every site was triaged:
  • rocksdb, 13 added:
    • stringappend2.cc's it->data() and it->size() on a std::deque<Slice>::const_iterator now reach Slice;
    • 11 calls in data_structure.h on auto it = findPosition(i) reach Interval::start, end and has_end. findPosition returns std::vector<Interval<Slice, Comparator>>::iterator. They land on the primary template's methods, the same targets main already gives . calls on that type in the same function.
  • rocksdb, 1 removed: itr->c_str() on a std::vector<std::string>::const_iterator had linked to gtest's FilePath::c_str. It now links to nothing, which is correct for a std::string element.
  • fmt and rocksdb, 1 metadata-only change each: the vendored gtest's property_with_matching_key->SetValue on a std::vector<TestProperty>::iterator keeps its target, with confidence up from 0.7/0.65 to 0.9.
  • leveldb, redis and protobuf: byte-identical. protobuf's iterator calls are on messages generated into descriptor.pb.h, which is 1.04 MB, just over the 1 MiB source-file cap. So it's indexed with no nodes, and those calls go on as before.
• Full suite: 7,139 passed and 1 failed. The failure was daemon-pid-reuse.test.ts (MCP daemon takeover after SIGKILL), a process-timing test unrelated to resolution; it passed twice when rerun alone.

Remaining, out of scope: iterators that are auto locals initialized by a library container's begin()/find() (about 80 it->m() sites across rocksdb and protobuf), and (*it)->m() / (*it).m() dereferences (about 30). Each needs its own typing step.

🤖 Generated with Claude Code