fix(bench): Java oracle miscounts package-private test methods as call sites

`ground_truth.py`'s Java "method" definition pattern required an explicit
`public`/`private`/`protected` modifier -- so a package-private method (the
standard JUnit 5 convention for test methods, e.g. `void testFoo() { ... }`
with no modifier at all) was never recognized as a *definition*. A JUnit
test named after the production method it exercises (e.g. `void
initUpdateOwnerForm() throws Exception` testing `OwnerController.
initUpdateOwnerForm()`) fell through and got miscounted as a real *call
site* of the production method.

Found while building B15 (cross-language competitor benchmark): on
spring-petclinic, CALM, CodeGraph, and Ctxo all scored 0/1 "missing" a call
to `initUpdateOwnerForm()` that was never real -- the oracle's sole "hit"
was the test method's own declaration line. Verified live: with the
modifier now optional (same as the pre-existing class/interface patterns
in this same file already treat it), `git_grep_call_sites` correctly
returns 0 sites for that symbol.

Shared by B7/B12/B13/B15 (all import `ground_truth.py`), so this closes the
same false-positive class for every consumer, not just B15.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Author: Your Name

benchmarks/b12_tier1_tier2_tool_correctness/ground_truth.py

@@ -60,7 +60,20 @@ class Definition:
         (r"^\s*(?:public|private|protected)?\s*(?:static\s+)?(?:final\s+)?(?:abstract\s+)?class\s+(\w+)", "class"),
         (r"^\s*(?:public|private|protected)?\s*(?:static\s+)?(?:final\s+)?interface\s+(\w+)", "interface"),
         (
-            r"^\s*(?:public|private|protected)\s+(?:static\s+)?(?:final\s+)?"
+            # 2026-08-18 fix: the visibility modifier used to be MANDATORY
+            # (`(?:public|private|protected)\s+`), so a package-private
+            # method -- the standard JUnit 5 convention for test methods,
+            # e.g. `void initUpdateOwnerForm() throws Exception {` with no
+            # modifier at all -- was never recognized as a definition by
+            # `_looks_like_a_definition`, and fell through to be counted as
+            # a real CALL SITE of any production method sharing its name.
+            # Verified live on spring-petclinic: a fresh B15 run scored
+            # CALM/CodeGraph/Ctxo 0/1 "missing" OwnerController's
+            # `initUpdateOwnerForm()` -- the sole "oracle file" was that
+            # test method's OWN declaration line, not a real call anywhere.
+            # Now optional, same as the class/interface patterns above
+            # already treat it.
+            r"^\s*(?:(?:public|private|protected)\s+)?(?:static\s+)?(?:final\s+)?"
             r"[\w<>\[\],\s]+?\s+(\w+)\s*\([^;{]*\)\s*(?:throws\s+[\w,\s]+)?\{",
             "method",
         ),