Metal/CUDA GPU qualifiers (kernel, __global__, device, ...) mint sourceless god nodes #4550

Version

v8 @ 0.9.92 (86c57be)

Problem

Metal (.metal) and CUDA (.cu/.cuh) are C++-derived GPU languages and route through the C++ extractor. Their address-space / execution-space / function qualifiers are not part of the C++ grammar, so tree-sitter-cpp parses a leading qualifier as the declaration's type_identifier (the real type is pushed into an ERROR node). _cpp_collect_type_refs then emits the qualifier as a type reference, and each one becomes a sourceless, corpus-shared god node that every shader/kernel references.

This violates the provenance invariant (a node with no source_file), and the hubs pollute god-node rankings in analyze.

Reproduce

Extract the shipped fixtures:

from graphify.extract import extract_cpp
r = extract_cpp("tests/fixtures/sample.metal")
print(sorted({n["label"] for n in r["nodes"] if not n.get("source_file")}))
# -> ['constant', 'device', 'kernel', 'uint']
r = extract_cpp("tests/fixtures/sample.cu")
print(sorted({n["label"] for n in r["nodes"] if not n.get("source_file")}))
# -> ['__device__', '__global__', ...]  (ids normalize to global/device/host)

Expected vs actual

• Expected: kernel, device, constant, __global__, __device__, __host__, ... are language keywords, not corpus symbols, and should not become nodes. The genuine struct Vec3 should still resolve.
• Actual: each qualifier is a sourceless god node with a false references edge.

Notes

These words are legal identifiers in plain C++/CUDA (e.g. a user type named device), so any filter must be keyed by the GPU-dialect file extension, not applied to .cpp/.cc/.h.