[colbymchenry]

Summary

Name matching weighed a candidate by its raw language tag, so calls between C, C++ and Objective-C files were treated as crossing languages: a call whose name had more than one definition in the other two languages linked to nothing, and the one definition of a name was a 0.5 "cross-language" guess. Now:

• the one candidate of a name is judged by language family (TypeScript/TSX/JavaScript, Java/Kotlin/Scala, C/C++/Objective-C/Swift, …), so a .tsx component calling the only .ts function of its name, or a C test calling the C API that a .cc file defines, is a 0.9 match instead of 0.5;
• C, C++ and Objective-C are ranked as one language when a name has several candidates: their files share one table of names, and a .h file is C, C++ or Objective-C only as far as sniffing its text can tell (Dear ImGui's C++ imgui.h reads as C). An Objective-C .m file's _glfwPlatformSetTls(…), which glfw defines in both posix_thread.c and win32_thread.c, now links to the nearer one; before, both scored −80 and the call linked to nothing.

Reimplements #1662 by @ferrine. Its headline (dropping cross-family bare-name calls) landed with #2032; this is the raw-tag-versus-family comparison it also pointed at.

Root cause

matchByExactName's single-candidate confidence, findBestMatch's same-language skip and its +50/−80 language score, and matchFuzzy's single-candidate confidence all compared candidate.language === ref.language. They predate sameLanguageFamily, which the cross-family gate already uses.

Fix, and why the ranking change is narrower than the diagnosis

• Single-candidate confidence (exact and fuzzy) compares families.
• findBestMatch and the weighed set beside it compare through sameRankingLanguage: the same language, or both of C, C++ and Objective-C, with the very language worth one point, which only breaks a tie (a directory nearer counts 15).

Ranking every family by family, as the triage proposed, was measured and rejected. With it, excalidraw's this.app.store.scheduleCapture() calls in the App.*.ts files moved from Store::scheduleCapture (right) to App::scheduleCapture in the App.tsx beside them (wrong), a JSON import was linked to a constant in a build script, and okhttp gained 83 Java-to-Kotlin new X.Builder() / X.Factory guesses onto whichever Builder was nearest (new OkHttpClient.Builder() → MockResponse.Builder). In those families files reach each other's names through imports, and the raw tag had been keeping a name-only pick inside one language. The new test pins the excalidraw shape.

Validation

• New __tests__/name-match-language-family.test.ts (8 tests, default and wasm): TSX → the one TS function at 0.9; a .ts call still ranked to the .ts class over a same-named .tsx method beside it; Objective-C → C (the one function at 0.9, the nearer of two); C++ → C (the nearer of two). On the base main 6 fail (the ranking pin passes on both); all 8 pass with the fix.

• A/B, base main vs fix (wasm on both arms; resolution only; repo commits in parentheses):

  Repo	Languages	Edges added / removed	Confidence 0.5 → 0.9, same target	Triage
  glfw (92dcf4ce)	C, Objective-C	4 / 0	126	right: nsgl_context.m → posix_thread.c's _glfwPlatformSetTls/GetTls and macos_time.c's timer functions, the macOS build's
  imgui (f88ec043)	C++, Objective-C++, headers read as C	93 / 0	446 (+6 fuzzy 0.3 → exact 0.4)	right: g.ColorStack.back() / pop_back() / erase() / find() → ImVector's methods in imgui.h (85), stbrp_pack_rects / stbrp_init_target in imstb_rectpack.h (2), ImWchar(…) conversions to its typedef (6, as a same-language C++ call already links)
  leveldb (7ee830d0)	C++, C	0 / 0	97	e.g. db/c_test.c → leveldb_get in db/c.cc
  jq (49fbca12)	C, C++	0 / 0	30	e.g. the C++ fuzzers → jq_init in src/execute.c
  excalidraw (d3a9c2a1)	TS, TSX	0 / 0	43 (+1 framework 0.8 → exact 0.9, +1 fuzzy 0.3 → 0.5)	unique names across .ts/.tsx, e.g. CanvasRenderingContext2D, App
  okhttp (c3179301)	Java, Kotlin	0 / 0	9	e.g. Java tests → Kotlin's OkHttpClient, OkHttpClientTestRule
  flask (d086db85), gin (0f09c3a9)	—	byte-identical	—	controls

• Main moved under this branch three times in name-matcher.ts (fix(python): a parameter or local named like an imported module is not that module #3046 Python, fix(ruby): calls to !/? methods and calls on what .new made link their method #3049 Ruby, fix(jvm): a dotted supertype names its nested type, and a Kotlin object expression is an anonymous class #3051 JVM). After each rebase the A/B on glfw, imgui, excalidraw, okhttp and flask was re-run on the kernel path against that main (kernel rebuilt), and gave the same deltas each time; the test fails 6 of 8 with the resolver of the fix(ruby): calls to !/? methods and calls on what .new made link their method #3049 and fix(jvm): a dotted supertype names its nested type, and a Kotlin object expression is an anonymous class #3051 mains (the final base's resolver is fix(jvm): a dotted supertype names its nested type, and a Kotlin object expression is an anonymous class #3051's) and passes 8 of 8 with the fix on every base.

• Index time, kernel path, median of 3 interleaved: imgui 7.9 s main, 7.7 s fix; excalidraw 9.2 s main, 9.4 s fix (noise).

• 24 related test files (cross-language, React Native and Expo bridges, Fabric, frameworks, mixed TS/JSX, Objective-C, resolution, PR19): 746 passed, 0 failed.

• Full suite on this exact commit (base main 19b74d1, kernel built, CODEGRAPH_KERNEL_EXPECT=1, CODEGRAPH_NO_DAEMON unset, 4 workers): 801 files, 9,520 passed, 0 failed, 86 skipped.

Left out

• imgui.h being read as C is a header-detection question of its own; its new links land on ImVector's inline methods as the C parse extracts them.
• TS/TSX and Java/Kotlin multi-candidate ranking keeps the raw tag (see above), so T6's pr1662-rank fixture (a .tsx call with a .ts sibling and a far .tsx namesake) still picks the .tsx one.
• Edge confidence is read today by the viewer's "uncertain" folding (under 0.6) and by the pick between strategies; in this A/B the latter changed one excalidraw edge's label (framework 0.8 → exact-match 0.9, same target).

🤖 Generated with Claude Code