Dart: capitalized ordinary function used as local-var initializer can still misroute/drop a receiver call (residual risk after #2474)

[carlos-alm]

Context

PR #2567 (fixing #2474) added typeMap seeding for var svc = UserService(repo); in Dart's extractor (handleDartLocalVarTypeMap / handle_dart_local_var_type_map), gated on the callee being capitalized to avoid misclassifying an ordinary lowercase factory function (var svc = makeService();) as a constructor call.

A Greptile review on that PR correctly pointed out the capitalization gate is not sufficient: Dart does not require a class name and an ordinary function name to differ in capitalization — a legally-named uppercase top-level function (OrderService MakeOrderService() { ... }) is syntactically indistinguishable from a constructor call at this position, and gets wrongly seeded as if the local's type were the literal function name.

What I verified empirically (not just reasoned about)

Added a dual-engine integration test reproducing this exact shape:

class OrderService {
  void placeOrder() {}
}

OrderService MakeOrderService() {
  return OrderService();
}

void placeAnOrder() {
  var order = MakeOrderService();
  order.placeOrder();
}

Result on BOTH engines: the placeAnOrder -> OrderService.placeOrder call edge is dropped entirely (not misrouted to a wrong node in this single-candidate case, but genuinely missing).

Root cause, traced through both resolvers:

• WASM/TS: resolveByReceiver (src/domain/graph/resolver/strategy.ts) branches on if (typeName) { ...only tries the (possibly wrong) type's qualified method / prototype alias... } else { ...tries resolveViaDirectQualifiedMethod using the receiver's OWN bare name... }. Once any typeMap entry exists for the receiver — right or wrong — the else branch (the receiver-name-based fallback) is never attempted at all.
• Native/Rust: resolve_call_targets_core (crates/codegraph-core/src/domain/graph/builder/stages/build_edges.rs) has the identical structural property — step "3.5. Direct qualified method lookup ... when the receiver is a class name with no typeMap entry" is guarded by if type_lookup.is_none() && inline_new_type.is_none(), so it's skipped whenever any type_lookup entry exists, right or wrong.

This is a pre-existing, language-agnostic property of the shared resolver, not something #2474 introduced — any receiver-type guess (from any language's extractor) that turns out to have zero real methods forecloses the untyped/direct-qualified fallback tier for that call, for every language that funnels through this shared cascade. #2474's Dart fix merely creates one more situation (an uppercase ordinary function) where a wrong guess can occur.

I did NOT attempt to fix the shared resolver cascade in #2567 — merging the type-aware and no-type-fallback branches (so a failed type-aware lookup falls through to try the receiver's-own-name qualified lookup too) is a resolver-architecture change with blast radius across every language using this cascade, not a Dart-specific fix, and deserves its own dedicated investigation rather than being folded into a single-language typing bug fix.

Two independent, worth-separating problems

1. Shared resolver: a receiver-type guess (right OR wrong) unconditionally forecloses the untyped/direct-qualified fallback tier, in both resolveByReceiver (TS) and resolve_call_targets_core (Rust). Fixing this (e.g., falling through to the receiver-name-based lookup when the type-aware lookup finds literally nothing) would help many languages, not just Dart, and reduce (though not eliminate) the practical impact of any heuristic-based type guess anywhere in the codebase.
2. Dart-specific: src/extractors/dart.ts's walkDartNode is a single combined tree-walk (unlike the Rust extractor's two-pass match_dart_node + match_dart_type_map design, and unlike javascript.ts's equivalent two-pass convention). This means a same-file negative check ("is this capitalized callee a known ordinary function already extracted into ctx.definitions, not a class?") would be declaration-order-dependent in the TS/WASM engine, while reliable in the Rust/native engine's two-pass design — an engine-parity hazard if implemented only on one side. Refactoring walkDartNode into a two-pass design to make this check reliable in both engines is a bigger, dedicated architectural change.

Suggested follow-up work (either or both)

• Investigate whether resolveByReceiver/resolve_call_targets_core can safely fall through to the receiver-name-based/global fallback tier when the type-aware tier finds zero real matches, across all languages that use this cascade (careful: needs its own precision/recall validation against the resolution-benchmark fixture suite, since loosening this could introduce new false-positive edges for other languages).
• Refactor src/extractors/dart.ts's walkDartNode into the same two-pass (match_dart_node-then-match_dart_type_map-equivalent) design already used by javascript.ts and mirrored in dart.rs, enabling a reliable same-file "is this name already a known ordinary function?" negative check for the local-var-constructor-call heuristic (and any future Dart heuristics that need the same kind of same-file cross-reference).

Current mitigation shipped in #2567

The capitalization gate (only seed when the callee is capitalized) meaningfully narrows — but does not eliminate — the false-positive rate, matching this codebase's own existing precedent for the identical class of ambiguity in javascript.ts's Foo.create() factory-method heuristic (confidence 0.7, also gated on capitalization alone, also carrying the same theoretical coincidental-collision risk). #2567 lowers this Dart heuristic's confidence to the same 0.7 tier for consistency.