[colbymchenry]

What

When several files declare a name a C or C++ test uses, the test now links to one its translation unit declares, when there is one:

• a definition in the test's own file;
• one in a file it includes, directly or through other includes;
• one in the source file that implements such a header (the same-stem file that includes it), unless that source keeps it to itself: a static, an unnamed namespace, or a class only it declares (fix(cpp): a call or type name never links to what only another source file can see #2538's isCppTranslationUnitLocal).

Before, those candidates were ranked by directory and index order, so a test took whichever namesake came first.

• protobuf's C# generator tests called test_util2.h's TestUtil::TestSourceDir(), a test suite none of them includes. The TestSourceDir() they call is the one googletest.h declares. This is the issue's own example.
• protobuf's upb and hpb tests left upb_Arena_New() and the rest of upb's C API unlinked (about 2,300 calls). The copies bundled for PHP and Ruby (php-upb.h, ruby-upb.h) won the ranking and were then rejected.
• rocksdb's db/c_test.c left rocksdb_open() and the rest of the C API in db/c.cc unlinked (about 1,400 calls).
• rocksdb's histogram test sent clock->SleepForMicroseconds() to emulated_clock.h's EmulatedSystemClock, a header it doesn't include.
• llm.c's test_gpt2_fp32.cu includes train_gpt2_fp32.cu directly, yet its gpt2_update(…) and cublasCheck(…) went to the other training programs.

Scope, and what it deliberately doesn't do

• Test files only. Production references keep fix(c,cpp): a file named like a test is visible to the code that includes it #2421's rule unchanged: a test suite is visible to them only through includes, checked after ranking.
• Where it applies: the exact-name strategy, typed receivers (resolveMethodOnType), and matchMethodCall's class-name and receiver-word strategies. The qualified-name strategy is unchanged (see the qualified-name note below).
• Never guesses outside the unit. When the unit declares none of the candidates, they stay as they were. A test whose includes don't resolve, like jemalloc's tests including a generated test/jemalloc_test.h, keeps the one helper there is. I measured a "drop out-of-TU test suites first" step and a veto variant. Both removed right edges or added wrong ones where main had correctly linked nothing:
  • vendored gtest's amalgamated gmock-gtest-all.cc defines what gtest.h declares under another name;
  • breakpad's EXPECT_CALL(memory, GetBase()) reached a production GetBase;
  • rocksdb's table_factory->GetOptions<…>() reached DBImpl::GetOptions.
• A name written alone (no receiver, no qualifier) is narrowed only to what C++ finds for it outside a class: a function, a namespace-scope type, or a constructor. Never another class's method or nested type: gmock's Value(result, …) matcher is no RepeatedFieldAccessor::Value, and breakpad's EXPECT_CALL(handler, Entry(…)) is no CallFrameInfo::Entry struct. A construction narrows to types and constructors only (std::string buffer(n, c) is no enum's string). A call the extractor recorded without the receiver it is written on (result.columns()[0].value()) is left as it was, read with fix(cpp): a call a method makes with no receiver links to its class's own or inherited method #2479's cppImplicitReceiver.
• C++ method calls to same-named classes in different files all resolve to the first definition #1079 is unchanged: the call site's own file is in its unit and still ranks first.

Re-verified on current main first

Most of the issue's original symptom was already fixed on main 3695818 by #2538 (TU-local classes), #2537 (TEST_F bodies reach the fixture) and #2479. Measured with the issue's own script, test-noinclude.mjs <db> '_test\.(cc|cpp|h)$'. It counts edges from test files into test files the source's translation unit doesn't include:

	at 1fb8691 (issue)	main 3695818	this branch
facebook/rocksdb	23,003	4	4
protocolbuffers/protobuf	48	4	3
google/leveldb	13	0	0
chromium/breakpad	1	0	0
google/glog	0	0	0
protobuf, _unittest\.(cc|cpp|h)$ sources	(18 with #2434)	20	14

rocksdb's 4 are synthesized base→override links from point_lock_manager_test.h into the tests that derive from it, which is by design.

The 14 protobuf unittest edges left are of two kinds:

• 10 constructions of generated messages (TestAllTypes, TestMessageSetExtension1) whose only in-repo namesakes sit in wire_format_unittest.h, a header these tests don't include (9 of them in an unnamed namespace);
• 4 TestUtil::ExpectClear calls taken by the qualified-name boundary bug (see the notes).

What this PR fixes is the rest:

• test-suite headers a test doesn't include, the TestSourceDir case;
• declarations in the test's own unit losing to namesakes outside it, which main left unlinked or wrong.

A/B (main 3695818 vs this branch), site-keyed edge diffs

Full indexes of 13 repos with the same kernel. A site is (kind, source node, line, col, refName).

Repo	Added	Retargeted	Removed	Confidence only (same target)
facebook/rocksdb	1,442	47	12	2,519
protocolbuffers/protobuf	2,316	55	67	546
fmtlib/fmt	33	40	0	128
chromium/breakpad	13	14	0	135
google/glog	22	0	0	0
google/leveldb	10	0	0	103
karpathy/llm.c	0	6	0	43
redis/redis	0	0	0	248
jqlang/jq, tokio, excalidraw, okhttp, typeorm	byte-identical

Triage. Every (before → after target) group was read. The added sites, by group:

• C API calls into the defining files: rocksdb rocksdb_* → db/c.cc; protobuf upb_* → upb's own .h/.c, not the PHP/Ruby bundles.
• glog's SetLogDestination(), FlushLogFiles() and LogToStderr() → logging.cc.
• leveldb's crc32c_test.cc Value(…) → crc32c::Value.
• fmt's format(…) → fmt::format, test_string(…) → util.h's alias, open_buffered_file() → test/util.cc.
• breakpad's WriteFile()/ReadFile() → common/tests/file_utils.cc.
• rocksdb's NewWritableFile(fs, …) → read_write_util.cc, and db_->Get(…)/wp_db->ReleaseSnapshot(…) → DB::.

Wrong among the 3,836 added: 11.

• protobuf: 7 arena_message1->GetArena() on a generated message → FieldWithArena::GetArena (receiver-word guess).
• rocksdb: 3 get<0>(GetParam()) → vendored gtest's std::tr1::get.
• fmt: 1 std::string_view construction → fmt's string_view alias.

Retargeted, 162 sites. About 145 are right or equivalent:

• right: TestSourceDir ×6; clock->SleepForMicroseconds → SystemClock ×12; CreateFile/Info/Log/ValidateOptions/Roundup/FindFile → their free functions; VisitFields, Prefetch<…>, ValidateEnum, the json:: stream functions; llm.c's included-source functions ×6; breakpad's per-platform CrashGenerationServer/ExceptionHandler and WriteMinidump → the free function; range_map->StoreRange → RangeMap;
• equivalent: a declaration in an included header in place of its out-of-line definition (DB::Put, Matcher::Matcher).

The rest of the retargets:

• wrong → wrong: rocksdb swap(…) ×1, the 4 trie_iter->Seek/Next on auto x = make_iter(…), protobuf FileDescriptorProto::descriptor() ×2, FieldOptions::default_instance() ×2;
• unclear: protobuf info.Clear() on a generic lambda's auto ×2;
• right → wrong: protobuf Varint(127) ×1 in upb's conformance harness, whose upb/wire/test_util/wire_message.h include doesn't resolve.

Removed, 79 sites. All were wrong receiver-word guesses: rocksdb result.size() → gtest's TestPartResultArray::size ×8 and 4 std_iter->Seek/Next guesses; protobuf FileDescriptorProto::descriptor()-style calls into java_features.pb.h / cpp_file_options.pb.h ×67.

Confidence only: the same target, now the only candidate the test's unit declares, so exact-match 0.4/0.7 → 0.9. Also 2 protobuf instance-method 0.8 → 0.65 with the same target.

Cost

Full indexes, interleaved main/branch, 3 rounds, same kernel:

• rocksdb: median 51.2s on main vs 48.3s here;
• protobuf: median 96.3s vs 100.5s.

The ranges overlap heavily (protobuf branch 80–115s, main 93–104s). The check runs only for a test's reference with two or more candidates, and reuses the memoized include graph (cppIncluders), the same-stem header list and the memoized test-path checks.

Tests

__tests__/cpp-test-translation-unit.test.ts, 5 cases. On main, 3 fail:

• the issue's TestSourceDir case;
• the upb copy;
• the rocksdb clock guess.

Two guards pass on both:

• C++ method calls to same-named classes in different files all resolve to the first definition #1079, the test's own file first;
• a test whose includes don't resolve keeps its one candidate.

Full suite on 08b8361 (this change on main 3695818), with the kernel built, CODEGRAPH_KERNEL_EXPECT=1 and --maxWorkers=4 --minWorkers=1: 7,481 passed, 1 failed, 80 skipped.

• The one failure was proxy-shutdown.test.ts's stopDaemonAt reports a gracefully stopping daemon as still-running, and the proxy fallback exits before its engine stops #2311 shutdown-timing test, under load. It passes alone (2/2).
• After syncing main bf6995e (5 commits, all JS/TS- or React-Router-only, no C/C++ path), tsc, this PR's test, cpp-included-test-suite and the 5 new test files from main pass (26/26).

Notes

• Qualified names. matchByQualifiedName's partial match isn't anchored at a :: boundary, so DB::Open(…) also matches BlobDB::Open and TransactionDB::Open, and Status::OK() matches IOStatus::OK. Narrowing those candidates to the unit retargeted 16 rocksdb calls right and 3 wrong, which is an artifact of the boundary bug. So this PR leaves that strategy alone. Branch claude/strange-greider-bd9f76 holds a boundary fix that never became a PR.
• Follow-up. The same "prefer what the unit declares" rule would help production C/C++ code too: production upb callers hit the same PHP/Ruby duplicates. It needs its own A/B. An earlier experiment that dropped out-of-TU test suites from production candidates before ranking added mostly wrong edges, so I'm filing it separately.

🤖 Generated with Claude Code