fix(extract): gate the #2551 partial-parse warning on plausible symbol loss (#2610, #2599)

The 0.9.37 #2551 warning fired on valid TypeScript/TSX: tree-sitter-typescript
sets root.has_error on tiny fully-recovered errors (a `&` in a JSX string
attribute; a semicolon-less `in_*` interface member) that still extract every
symbol. The warning now fires only when recovery plausibly cost symbols — the
file yielded <=1 node, or a materialized ERROR region spans more than one line —
so the genuine Kotlin one-line-body (#2551) and Luau (#2520) cases still warn
while valid TS goes silent. Warning also prints a root-relative path.

Thanks @Sid-AutoWisdom and @atlasplatformu-ai.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

graphify/extract.py

@@ -5471,12 +5471,24 @@ def extract(
     # naming the first error line so the user can find the construct.
     _syntax_error_files: list[tuple[str, int | None]] = []
     for i, _p in enumerate(paths):
-        _pe = (per_file[i] or {}).get("parse_errors")
-        if _pe:
-            _syntax_error_files.append((str(_p), _pe.get("first_error_line")))
+        _res = per_file[i] or {}
+        _pe = _res.get("parse_errors")
+        if not _pe:
+            continue
+        # #2610/#2599: gate the #2551 warning on plausible symbol loss.
+        # tree-sitter-typescript sets has_error on tiny fully-recovered errors
+        # (a `&` in a JSX string attr; a semicolon-less `in_*` interface
+        # member) that extract completely — stay silent. Warn only when
+        # nothing beyond the file node extracted, or an ERROR region
+        # dissolved multiple lines (the genuine #2551 Kotlin one-line-body /
+        # #2520 Luau case). `multiline_error` is absent from pre-fix cached
+        # results, so those fall back to the file-node-only arm.
+        if len(_res.get("nodes", [])) <= 1 or _pe.get("multiline_error"):
+            _rel = os.path.relpath(str(_p), str(root)).replace("\\", "/")
+            _syntax_error_files.append((_rel, _pe.get("first_error_line")))
     if _syntax_error_files:
         _shown = ", ".join(
-            f"{Path(x).name} (first error at line {ln})" if ln else Path(x).name
+            f"{x} (first error at line {ln})" if ln else x
             for x, ln in _syntax_error_files[:5]
         )
         _more = (

graphify/extractors/engine.py

@@ -2344,6 +2344,19 @@ def _first_parse_error_line(root) -> int:
         node = child
 
 
+def _has_multiline_error(root) -> bool:
+    """True if any materialized ERROR node spans more than one line (a
+    recovery region large enough to plausibly drop symbols, vs a tiny
+    single-line recovery that extracts completely)."""
+    stack = [root]
+    while stack:
+        n = stack.pop()
+        if n.type == "ERROR" and n.end_point[0] > n.start_point[0]:
+            return True
+        stack.extend(c for c in n.children if c.has_error)
+    return False
+
+
 def _read_csharp_type_name(node, source: bytes) -> tuple[str, bool, str] | None:
     """Resolve a C# type name, whether it was qualified, and its qualifier prefix."""
     if node is None:
@@ -5255,7 +5268,10 @@ def _extract_generic(
     # error's line so extract() can warn instead of reporting silent success.
     # Rides on the result dict, so it survives the per-file AST cache.
     if root.has_error:
-        result["parse_errors"] = {"first_error_line": _first_parse_error_line(root)}
+        result["parse_errors"] = {
+            "first_error_line": _first_parse_error_line(root),
+            "multiline_error": _has_multiline_error(root),
+        }
     # Kotlin (#2526/#2550): the declared package qualifies every node in the
     # file; the import-target and qualified-call resolvers key their per-package
     # symbol indexes off it.

tests/test_kotlin_grammar.py

@@ -15,9 +15,12 @@ older forks the extractor was written against:
   all-identifier chains into a `qualified_prefix` resolved against the
   declared packages (exactly-one-candidate guarded).
 * #2551 — the grammar rejects one-line `class C { val x }` bodies; consecutive
-  one-liners can dissolve the whole file's parse. Graphify now warns on any
-  file extracted through ERROR recovery (language-agnostic, also #2520) and
-  keeps class linkage for declarations recovered inside an ERROR span.
+  one-liners can dissolve the whole file's parse. Graphify warns on a file
+  extracted through ERROR recovery (language-agnostic, also #2520) and keeps
+  class linkage for declarations recovered inside an ERROR span. Since
+  #2610/#2599 the warning fires only on PLAUSIBLE symbol loss (file-node-only
+  result or a multiline ERROR region) — tiny fully-recovered errors that
+  extract completely stay silent.
 """
 from __future__ import annotations
 
@@ -272,7 +275,10 @@ def test_kotlin_partial_parse_warns_with_file_and_line(tmp_path, capsys):
 
 def test_kotlin_one_line_class_with_fun_still_extracts(tmp_path, capsys):
     # `class VM { fun f() = 1 }` trips has_error but recovers structurally:
-    # everything must extract, and the warning names the file.
+    # everything must extract. #2610: since the recovery is zero-width and
+    # every symbol is present, the corrected gate (warn only on plausible
+    # symbol loss — file-node-only or a multiline ERROR region) stays SILENT;
+    # the old has_error-gated warning here was a false positive.
     r = _extract(tmp_path, {
         "VM.kt": (
             "class VM { fun f() = 1 }\n"
@@ -283,7 +289,7 @@ def test_kotlin_one_line_class_with_fun_still_extracts(tmp_path, capsys):
     f = _find(r, ".f()")
     _find(r, "After()")  # present
     assert (vm, f) in _edges(r, "method")
-    assert "VM.kt" in capsys.readouterr().err
+    assert "VM.kt" not in capsys.readouterr().err
 
 
 def test_kotlin_one_line_class_keeps_field_reference(tmp_path):

tests/test_ts_parse_warning.py

@@ -0,0 +1,120 @@
+"""#2610/#2599: the #2551 partial-parse warning must not fire on VALID TS/TSX.
+
+tree-sitter-typescript (0.23.x) sets ``has_error`` on tiny, fully-recovered
+errors in code ``tsc`` accepts — a ``&`` inside a JSX string attribute, or a
+semicolon-less interface member named ``in_*``/``instanceof_*`` — while still
+extracting every symbol. Gating the warning on bare ``has_error`` (0.9.37,
+#2551) therefore flagged valid files as "syntax errors". The corrected gate
+warns only on PLAUSIBLE symbol loss: nothing beyond the file node extracted,
+or an ERROR region spanning multiple lines.
+"""
+from __future__ import annotations
+
+import os
+import re
+from pathlib import Path
+
+from graphify.extract import extract
+
+
+def _extract(tmp_path, files: dict[str, str]):
+    for name, body in files.items():
+        p = tmp_path / name
+        p.parent.mkdir(parents=True, exist_ok=True)
+        p.write_text(body)
+    old = os.getcwd()
+    try:
+        os.chdir(tmp_path)
+        r = extract([Path(n) for n in files],
+                    cache_root=tmp_path / ".cache", parallel=False)
+    finally:
+        os.chdir(old)
+    return r
+
+
+def _labels(r):
+    return {n["label"] for n in r["nodes"]}
+
+
+def _assert_silent(err):
+    assert "syntax errors" not in err
+    assert "partially extracted" not in err
+
+
+def test_tsx_amp_in_jsx_string_attr_is_silent(tmp_path, capsys):
+    # `&` inside a JSX string attribute trips a grammar lexer bug: a 4-byte
+    # ERROR node covering `&b=2`, has_error=True — but `el` extracts fine and
+    # tsc accepts the file. No warning.
+    r = _extract(tmp_path, {
+        "app.tsx": (
+            "declare const Comp: (props: { to: string }) => null;\n"
+            'export const el = <Comp to="/p?a=1&b=2" />;\n'
+        ),
+    })
+    assert "el" in _labels(r)
+    _assert_silent(capsys.readouterr().err)
+
+
+def test_ts_interface_member_named_in_prefix_is_silent(tmp_path, capsys):
+    # A semicolon-less interface member starting with `in` (`in_workshop`)
+    # yields a zero-width MISSING `}` + a 1-byte ERROR, has_error=True — but
+    # `I` extracts fine and tsc accepts the file. No warning.
+    r = _extract(tmp_path, {
+        "types.ts": (
+            "interface I { a: number\n"
+            " in_workshop: number }\n"
+        ),
+    })
+    assert "I" in _labels(r)
+    _assert_silent(capsys.readouterr().err)
+
+
+def test_ts_generics_as_and_jsx_logical_and_are_silent(tmp_path, capsys):
+    # Regression canary: constructs superficially near the lexer bugs —
+    # `foo<T>(x)`, an `as` cast, a `<T,>` arrow generic, `&&` in JSX — parse
+    # CLEAN (no has_error) and must never warn.
+    r = _extract(tmp_path, {
+        "generics.ts": (
+            "function foo<T>(x: T): T { return x }\n"
+            "const y = foo<number>(1);\n"
+            "const z = y as number;\n"
+        ),
+        "arrow.tsx": (
+            "const pick = <T,>(x: T) => x;\n"
+            "export const view = <div>{true && <span>hi</span>}</div>;\n"
+        ),
+    })
+    assert {"foo()", "pick()", "view"} <= _labels(r)
+    _assert_silent(capsys.readouterr().err)
+
+
+def test_ts_genuinely_broken_file_still_warns(tmp_path, capsys):
+    # `function f( {` dissolves the whole parse — nothing beyond the file
+    # node extracts. The warning must fire and name the file + first line.
+    _extract(tmp_path, {"broken.ts": "function f( {\n"})
+    err = capsys.readouterr().err
+    assert "syntax errors" in err
+    assert "broken.ts" in err
+    assert re.search(r"first error at line 1", err)
+
+
+def test_ts_midfile_breakage_warns_and_keeps_intact_functions(tmp_path, capsys):
+    # An unclosed brace between two valid functions produces a multiline
+    # ERROR region: the warning fires AND both intact functions extract.
+    r = _extract(tmp_path, {
+        "midfile.ts": (
+            "function alpha() { return 1 }\n"
+            "\n"
+            "function bad( {\n"
+            "  oops;\n"
+            "  more;\n"
+            "\n"
+            "function beta() { return 2 }\n"
+        ),
+    })
+    labels = _labels(r)
+    assert "alpha()" in labels
+    assert "beta()" in labels
+    err = capsys.readouterr().err
+    assert "syntax errors" in err
+    assert "midfile.ts" in err