fix(extract): don't bind TS/JS builtin static calls (Date.now(), …) to same-named user symbols (#1726)

[2loch-ness6]

What

Fixes #1726 — TS/JS static calls on ECMAScript built-in globals (Date.now(), Object.keys(), Math.max(), …) bind to a same-spelled user symbol, creating a false god-node.

Root cause

_resolve_typescript_member_calls treats a capitalized member-call receiver as a type name and matches it through _key(), which case-folds:

if receiver[:1].isupper():
    type_name = receiver            # "Date"
...
type_defs = type_def_nids.get(_key(type_name), [])   # _key("Date") == "date"

If the corpus has a lone user symbol whose name case-folds to the same key — e.g. a module-local const DATE (_key("DATE") == "date") — it is the unique match. Since the const has no now method, the resolver falls to relation = "references" and emits a references/call edge onto it. Date.now() (and friends) appear in nearly every file, so the one user symbol accumulates hundreds of cross-file edges.

Observed on a real 3,368-file repo: a non-exported const DATE became the #1 god-node — 472 edges, betweenness 0.435, bridging ~130 communities. Every source was a method calling new Date() / Date.now() in unrelated packages.

The call-edge pass (if callee in _LANGUAGE_BUILTIN_GLOBALS: continue) and the raw-call pass already guard against exactly this. The member-call resolver was the one path that didn't.

Fix

One guard, mirroring the existing ones — skip the receiver when it is a language builtin:

if type_name in _LANGUAGE_BUILTIN_GLOBALS:
    continue

Date is already in _LANGUAGE_BUILTIN_GLOBALS (alongside Object, Array, Math, JSON, Promise, …), so no list changes are needed.

Scope / safety

• Legitimate receiver-typed member calls are unaffected — the resolver's real job (constructor-injection type tables, this.repo.findById() → UserRepo.findById) still resolves. Covered by an added test.
• A user type genuinely named like a builtin (class Date) matched case-sensitively would have resolved before this branch anyway; a builtin receiver never legitimately targets an unrelated same-folded user symbol.

Tests

tests/test_case_sensitive_resolution.py:

• test_ts_builtin_static_call_does_not_resolve_to_user_symbol — reproduces the leak (fails without the guard, passes with it).
• test_ts_receiver_typed_member_call_still_resolves — pins that legitimate cross-file member-call resolution is preserved.

Full suite green locally (2848 passed; the only failures are pre-existing tree_sitter_hcl not installed Terraform tests, unrelated to this change).

[safishamsi]

Thanks for this — you independently arrived at the exact same fix. This shipped in 0.9.10 (67f4f83, #1726): the identical guard if type_name in _LANGUAGE_BUILTIN_GLOBALS: continue at the same spot in _resolve_typescript_member_calls, so Date.now() / Object.keys() / Math.max() etc. no longer bind cross-file to a same-spelled user symbol.

I verified your Date.now() static-call shape (typed param arming the type table) against current v8: the cross-file phantom is gone, while a legitimate same-file reference to a real const DATE is preserved. I also folded your static-call test shape into the regression suite (crediting you). Closing as already implemented — appreciate the independent confirmation and the extra test case.