Dart extractor emits source_location: None for every node — 0% on 3 repos, ~25k nodes unciteable (regex extractor hardcodes the field)

[santoshpy]

Summary

Every node the Dart extractor emits has source_location: None. Not "most" — all of
them, in every Dart file, because extractors/dart.py hardcodes the field at all three
node/edge construction sites. Dart graphs therefore cannot cite a line, and no
downstream consumer can verify a node against the file it claims to come from.

This is worth separating from #3307, which asks for source_location on doc/YAML/markdown
and states as its premise that "code already gets this for free from AST (line numbers
are known at parse time)". Dart is a code language with an extractor, and it gets 0%.

Minimal reproduction

Two structurally identical files, ten lines each, one graphify extract . --code-only
run:

// sample.dart
class Greeter {
  final String name;
  Greeter(this.name);

  String greet() {
    return 'hello $name';
  }
}

int add(int a, int b) => a + b;

# sample.py
class Greeter:
    def __init__(self, name):
        self.name = name

    def greet(self):
        return f"hello {self.name}"


def add(a, b):
    return a + b

Resulting nodes:

.dart: 5 nodes, 0 with source_location      .py: 5 nodes, 5 with source_location
  None  sample.dart                            L1  Greeter
  None  add                                    L5  .greet()
  None  Greeter                                L2  .__init__()
  None  greet                                  L9  add()
  None  name                                   L1  sample.py

Scale, on real repositories

Measured across nine repos (all --code-only, graphifyy 0.9.55):

repo	dominant language	declared nodes	with source_location
Plugins	Dart	286	0 (0%)
Board	Dart	531	0 (0%)
Editor	Dart	5,625	0 (0%)
App	Dart + Rust	31,958	6,961 (22% — exactly the Rust half)
Cloud	Rust	6,304	6,223 (99%)
Web	TS/TSX	10,143	10,134 (100%)
Auth	Go	3,655	3,583 (98%)
ops	Python	707	707 (100%)

~25,000 Dart nodes in one repo with no position at all.

Root cause

extractors/dart.py is regex-based (import re, 31 re.* calls, zero tree_sitter
references — its docstring says "using regex"), so there is no start_point to read,
and the field is written as a literal:

nodes.append({"id": file_nid, "label": path.name, "file_type": "code",
              "source_file": str(path), "source_location": None})          # L54

def add_node(nid, label, ftype="code", source_file=str(path)) -> None:
    if nid not in defined:
        nodes.append({"id": nid, "label": label, "file_type": ftype,
                      "source_file": source_file, "source_location": None})  # L61

edge = {..., "source_file": str(path), "source_location": None, ...}         # L67

add_node takes no line parameter, so callers such as add_node(class_nid, class_name)
(L145) have nowhere to pass one.

Suggested fix, and a trap in it

Each match already knows its offset, so the line is src[:m.start(1)].count("\n") + 1.
Two details matter, both verified against the file below:

1. Anchor on the capture group, not the match. The class pattern begins ^\s*, so
   m.start() sits before the preceding newlines and reports the line too early.
2. _comment_replace deletes comments (returns ""), so offsets in the cleaned
   text no longer correspond to the original file. Blanking a comment to an equal number
   of newlines keeps every later line correct.

/*
 * A multi-line header comment.
 * It spans several lines.
 */
class AfterComment {      // real line 5
  void method() {}
}

true line                            : 5
today (comments deleted, m.start(1)) : 2     <- off by the comment's height
newlines preserved + m.start(1)      : 5     <- correct

So: return "\n" * token.count("\n") instead of "" for comment tokens, thread a
line argument through add_node/add_edge, and compute it from m.start(1) at each
re.finditer site.

Why it matters beyond citations

We cross-validate graphs against the source they describe (file-coverage ledger, then
per-node line anchoring, then recall against an independent parser). For Rust, Go, TS
and Python the line anchor confirms a node really is the symbol it claims to be at the
line it claims. For Dart that check cannot run at all, so ~78% of our largest repo's
graph is unverifiable — it reports n/a rather than a pass, and a Dart-only repo can
never be more than partially validated.

Environment

graphifyy 0.9.55 (uv tool, Python 3.12.12), macOS 15 arm64. graphify extract --code-only.