TS/JS `new`/type references to builtin globals (e.g. `Date`) collapse onto same-named user symbols → false god node

[2loch-ness6]

Summary

TS/JS references to ECMAScript built-in globals (Date, and any other builtin whose casefolded name matches a user symbol) collapse onto a same-named user declaration, producing a huge false "god node". In a real 3,368-file corpus, a module-local, non-exported const DATE in one frontend file accumulated 469 false cross-file references edges (degree 472, betweenness 0.435, bridging ~130 communities) — every new Date() call site in the repo resolved to it.

Call edges are already guarded against this; the type/new-reference path is not.

Root cause

Two things combine:

1. normalize_id() casefolds (ids.py). The ECMAScript global Date and a user const DATE both normalize to date, so they are indistinguishable to the label/id index.

2. The reference-resolution path does not consult _LANGUAGE_BUILTIN_GLOBALS. Call edges do — every call-edge site guards on it, e.g.:

   # extract.py:16730
   if callee in _LANGUAGE_BUILTIN_GLOBALS:
       ...  # skip

   But ensure_named_node(name, line) (extract.py:3580, and its 6 sibling closures) has no such guard. When a new Date() / Date reference is walked, it creates a sourceless stub (id: date, label: Date) and draws a references edge (context="call", confidence=EXTRACTED, score 1.0). The corpus-level rewire/dedup step then collapses that stub onto the single real declaration whose norm-id matches — the user's const DATE.

Net: Date ∈ _LANGUAGE_BUILTIN_GLOBALS (so calls are skipped), yet the references/context="call" path still emits edges to a user symbol.

Proposed fix

Guard the reference path the same way call edges are guarded — the natural chokepoint is ensure_named_node: if name in _LANGUAGE_BUILTIN_GLOBALS, do not create a stub/edge (return a sentinel the callers skip, mirroring the existing if target_nid != func_nid skip pattern). Apply consistently to all ensure_named_node closures.

Optionally, as defense in depth, normalize _BUILTIN_NOISE_LABELS membership tests through normalize_id so casefold collisions (DATE vs Date) are still filtered from god-node reporting.

Environment

• graphifyy 0.9.9
• Python 3.12.3, Linux

[safishamsi]

Fixed on v8 (67f4f83), shipping in the next release.

Your diagnosis (casefold collision + no builtin guard on the reference path) was spot on. The actual emission site turned out to be _resolve_typescript_member_calls rather than ensure_named_node: for a builtin-typed receiver (x: Date; x.getTime()), the resolver matched the receiver type Date to a same-named user class DATE/const DATE by casefolded label (_key("Date") == "date") and bound the caller to it as a references[call] edge — which is how "every new Date() call site" collapsed onto the one const DATE.

The fix skips resolution when the receiver type is in _LANGUAGE_BUILTIN_GLOBALS, mirroring the guard the cross-file CALL resolver already applies (#726). It's a strict no-op for genuine user types, so legitimate constructor-injection member-call resolution (#1316) is unaffected. Verified the phantom edges are gone and the user DATE node drops back to degree 1 across new Date().method(), return-type, and var-decl-type shapes; regression tests added for both the fix and the no-op-on-user-types case.

Closing as fixed — please reopen if you spot any residual builtin→user collapse on another reference path (e.g. a language other than TS/JS).