Python constructor calls do not reach the initializer they run #4209

Python constructor calls stop at the class node, hiding the initializer's dependencies from change-impact traversal.

Reproduction on current v8

Verified on v0.9.80, commit 6478eb7, with Python 3.13 on Linux. Build these files with graphify update <fixture> --no-cluster:

# pkg/storage.py
def open_db(path):
    return path

class Store:
    def __init__(self, path):
        self.db = open_db(path)

# pkg/cache.py
from pkg.storage import Store

class CachedStore(Store):
    pass

# pkg/service.py
from pkg.cache import CachedStore

def build(path):
    return CachedStore(path)

# tests/test_service.py
from pkg.service import build

def test_build():
    build("test.db")

graphify affected "open_db()" --depth 4 --graph <fixture>/graphify-out/graph.json finds Store.__init__ but misses build() and test_build(). The call graph has build -> CachedStore and Store.__init__ -> open_db, with no call edge joining them.

Constructor fix

PR #4208 adds an inferred constructor-entry edge when a supported source binding and class chain identify the initializer. It preserves the class edge and every other existing node and edge. The fixture then reaches the initializer, build(), its importing test file, and test_build(). Incremental output after editing the caller matches a clean rebuild.

Fresh comparisons against 6478eb7 add 703 constructor edges across seven pinned repositories: Flask 3, requests 3, httpx 56, rich 506, langchain-core 100, yt-dlp 17 and FastAPI 18. All 703 agree with an independent Python ast derivation. All 586 that SCIP can judge agree on the called class, inheritance chain and initializer. Every pre-existing node and edge remains identical.

This is a restricted static subset. Decorators, metaclasses, custom allocation, multiple inheritance, ambiguous bindings, external bases and known mutation patterns add no edge. Cross-module dynamic mutation remains a documented limitation.

Why it matters for test impact

A separate runtime-trace study measured Flask and yt-dlp tests against pinned source revisions. Historical v8 graphs reached only 4 of 3,482 sampled Flask test/function pairs and 20 of 606 yt-dlp pairs. A broader prototype combining receiver typing, constructor entry and pytest fixture dependencies reached 1,699 and 358 respectively. Those are historical prototype results, not the results of this constructor PR or current v0.9.80.

Constructor entry is one missing link in that chain. Since the study, upstream has added self.attr.method() support in #4176 and annotated/constructor-bound local receiver support in #4198. Those changes are already present in the base used to verify this PR. Remaining opportunities include inherited methods, factory return types and pytest fixture dependency edges through the nearest conftest.py chain.

Related: #2041, #2860, #2417, #2586 and #1499.