feat(extraction): add Erlang language support (.erl/.hrl) (#635, #648) (#1165)

Vendored WhatsApp/tree-sitter-erlang 0.19 (the ELP grammar, ABI 14) with an
Erlang-shaped extractor: multi-clause/multi-arity functions merged into one
symbol, -spec signatures, records with fields, -type/-opaque aliases, -define
macros, -include/-include_lib file edges, and -export-driven visibility.

Modules wrap in a namespace so remote mod:fn(...) calls resolve through the
existing qualified-name matcher as mod::fn with zero resolver changes.
-behaviour declarations link to the behaviour module — gated to namespace
targets only (bare-name fallthrough linked -behaviour(supervisor) to an
unrelated macro constant on emqx). OTP indirection with static targets is
followed: spawn/apply/proc_lib/timer/rpc MFA-argument callees, and
gen_server:call/cast(?MODULE | ?SERVER) to the module's own
handle_call/handle_cast. Var-module dispatch and message sends stay
deliberately unlinked. codegraph_explore also normalizes Erlang-native query
spelling (mod:fn/3, init/2) so named symbols resolve as typed.

Benchmarked on cowboy (189 files), ejabberd (414), emqx (2,447): extraction
PASS on all three; with-codegraph arms reached 2/2/0 file Reads vs 10/5+/19
without, fastest on the largest repo.

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

src/extraction/grammars.ts

@@ -44,6 +44,7 @@ const WASM_GRAMMAR_FILES: Record<GrammarLanguage, string> = {
   cfquery: 'tree-sitter-cfquery.wasm',
   cobol: 'tree-sitter-cobol.wasm',
   vbnet: 'tree-sitter-vbnet.wasm',
+  erlang: 'tree-sitter-erlang.wasm',
 };
 
 /**
@@ -135,6 +136,10 @@ export const EXTENSION_MAP: Record<string, Language> = {
   // VB.NET: vendored grammar (patched govindbanura/tree-sitter-vbnet) — classes,
   // modules, interfaces, structures, properties, events, Handles clauses, LINQ.
   '.vb': 'vbnet',
+  // Erlang: modules (.erl) and header files (.hrl). Vendored WhatsApp/
+  // tree-sitter-erlang grammar (the ELP grammar).
+  '.erl': 'erlang',
+  '.hrl': 'erlang',
   // Spring config: `application.properties` / `application-*.properties`. Same
   // shape as the `.yml` variants — the YAML/properties extractor emits one node
   // per leaf key, and the Spring resolver links `@Value("${k}")` references.
@@ -253,7 +258,7 @@ export async function loadGrammarsForLanguages(languages: Language[]): Promise<v
       // `class Foo(...)` as an ERROR that swallows the whole class (#237); we
       // vendor the upstream ABI-15 tree-sitter-c-sharp 0.23.5 wasm, which parses
       // primary constructors natively.
-      const wasmPath = (lang === 'pascal' || lang === 'scala' || lang === 'lua' || lang === 'luau' || lang === 'csharp' || lang === 'r' || lang === 'cfml' || lang === 'cfscript' || lang === 'cfquery' || lang === 'cobol' || lang === 'vbnet')
+      const wasmPath = (lang === 'pascal' || lang === 'scala' || lang === 'lua' || lang === 'luau' || lang === 'csharp' || lang === 'r' || lang === 'cfml' || lang === 'cfscript' || lang === 'cfquery' || lang === 'cobol' || lang === 'vbnet' || lang === 'erlang')
         ? path.join(__dirname, 'wasm', wasmFile)
         : require.resolve(`tree-sitter-wasms/out/${wasmFile}`);
       const language = await WasmLanguage.load(wasmPath);
@@ -473,6 +478,7 @@ export function getLanguageDisplayName(language: Language): string {
     cfquery: 'CFQuery (SQL)',
     cobol: 'COBOL',
     vbnet: 'Visual Basic .NET',
+    erlang: 'Erlang',
     unknown: 'Unknown',
   };
   return names[language] || language;

src/extraction/languages/erlang.ts

@@ -0,0 +1,276 @@
+import type { Node as SyntaxNode } from 'web-tree-sitter';
+import { getNodeText, getChildByField, getPrecedingDocstring } from '../tree-sitter-helpers';
+import type { LanguageExtractor, ExtractorContext } from '../tree-sitter-types';
+
+// Node names follow the vendored WhatsApp/tree-sitter-erlang grammar (0.19,
+// ABI 14) — the grammar behind the Erlang Language Platform (ELP).
+//
+// Erlang is form-based, and three of its shapes don't fit the generic
+// extractor, so every symbol-bearing top-level form is dispatched through the
+// visitNode hook below instead:
+//   - a function's name lives on its CLAUSE, not the fun_decl, and the grammar
+//     emits one fun_decl PER CLAUSE — consecutive same-name fun_decl forms are
+//     merged into a single function node here;
+//   - type-position expressions (-spec/-type/-callback bodies, record field
+//     types) parse as `call` nodes, so descending into them would mint bogus
+//     call refs to type names (`pid()`, `term()`); the hook consumes those
+//     subtrees;
+//   - record_decl carries its fields as direct children (no body field), which
+//     the generic extractStruct would skip as a forward declaration.
+// Calls (local `f(X)`, remote `mod:f(X)`, `fun f/1` references, and record
+// usages) are handled by the erlang branch in extractCall — remote calls are
+// emitted as `mod::f`, which matches the qualifiedName the module namespace
+// produces (see packageTypes below), so cross-module resolution rides the
+// standard qualified-name matcher.
+
+/** Text of an atom with quoted-atom quotes stripped (`'EXIT'` → `EXIT`). */
+function atomText(node: SyntaxNode, source: string): string {
+  return getNodeText(node, source).replace(/^'([\s\S]*)'$/, '$1');
+}
+
+function collapseWs(text: string): string {
+  return text.replace(/\s+/g, ' ').trim();
+}
+
+// --- Per-file memos. Extraction is file-sequential within a worker, so a
+// single-entry memo keyed by filePath is safe (and resets naturally). ---
+
+/** Exported function names for the current file ('all' for -compile(export_all)). */
+let exportsFile = '';
+let exportsMemo: Set<string> | 'all' = new Set();
+
+/**
+ * Clause-merge state: the previous fun_decl's name and node id. A fun_decl
+ * whose clause repeats that name is a continuation clause (or a same-name
+ * different-arity definition — deliberately grouped under one node, the way
+ * overloads are elsewhere) and attaches to the existing node instead of
+ * creating a duplicate.
+ */
+let lastFnFile = '';
+let lastFnName = '';
+let lastFnId = '';
+
+function moduleExports(node: SyntaxNode, source: string, filePath: string): Set<string> | 'all' {
+  if (filePath === exportsFile) return exportsMemo;
+  let root: SyntaxNode = node;
+  while (root.parent) root = root.parent;
+  let result: Set<string> | 'all' = new Set<string>();
+  for (let i = 0; i < root.namedChildCount; i++) {
+    const form = root.namedChild(i);
+    if (!form) continue;
+    if (
+      form.type === 'compile_options_attribute' &&
+      getNodeText(form, source).includes('export_all')
+    ) {
+      result = 'all';
+      break;
+    }
+    if (form.type === 'export_attribute') {
+      for (const fa of form.namedChildren) {
+        if (fa.type !== 'fa') continue;
+        const fun = getChildByField(fa, 'fun');
+        if (fun) result.add(atomText(fun, source));
+      }
+    }
+  }
+  exportsFile = filePath;
+  exportsMemo = result;
+  return result;
+}
+
+/** The -spec directly above a function (comments may sit between), if it names it. */
+function precedingSpec(node: SyntaxNode, name: string, source: string): SyntaxNode | null {
+  let prev = node.previousNamedSibling;
+  while (prev && prev.type === 'comment') prev = prev.previousNamedSibling;
+  if (prev?.type === 'spec') {
+    const specFun = getChildByField(prev, 'fun');
+    if (specFun && atomText(specFun, source) === name) return prev;
+  }
+  return null;
+}
+
+/** `name(Args) when Guard` — the clause text up to the `->`. */
+function clauseHeader(clause: SyntaxNode, source: string): string | undefined {
+  const body = getChildByField(clause, 'body');
+  const end = body ? body.startIndex : clause.endIndex;
+  return collapseWs(source.substring(clause.startIndex, end)) || undefined;
+}
+
+function handleFunDecl(node: SyntaxNode, ctx: ExtractorContext): boolean {
+  const clauses = node.namedChildren.filter((c) => c.type === 'function_clause');
+  const first = clauses[0];
+  if (!first) return true; // macro-templated clause (`?M(...) -> ...`) — no static name
+  const nameNode = getChildByField(first, 'name');
+  if (!nameNode) return true;
+  const name = atomText(nameNode, ctx.source);
+  if (!name) return true;
+
+  // Continuation clause: extend the existing node's span and attribute this
+  // clause's calls to it.
+  if (ctx.filePath === lastFnFile && name === lastFnName && lastFnId) {
+    for (let i = ctx.nodes.length - 1; i >= 0; i--) {
+      const n = ctx.nodes[i];
+      if (n && n.id === lastFnId) {
+        if (node.endPosition.row + 1 > n.endLine) n.endLine = node.endPosition.row + 1;
+        break;
+      }
+    }
+    ctx.pushScope(lastFnId);
+    for (const clause of clauses) ctx.visitFunctionBody(clause, lastFnId);
+    ctx.popScope();
+    return true;
+  }
+
+  const spec = precedingSpec(node, name, ctx.source);
+  const exports = moduleExports(node, ctx.source, ctx.filePath);
+  const fn = ctx.createNode('function', name, node, {
+    docstring: getPrecedingDocstring(spec ?? node, ctx.source),
+    signature: spec
+      ? collapseWs(getNodeText(spec, ctx.source)).slice(0, 300)
+      : clauseHeader(first, ctx.source),
+    isExported: exports === 'all' || exports.has(name),
+  });
+  if (!fn) return true;
+  ctx.pushScope(fn.id);
+  // The whole clause is walked (not just the body) so record patterns in the
+  // arguments and guard calls contribute references too.
+  for (const clause of clauses) ctx.visitFunctionBody(clause, fn.id);
+  ctx.popScope();
+  lastFnFile = ctx.filePath;
+  lastFnName = name;
+  lastFnId = fn.id;
+  return true;
+}
+
+function handleRecordDecl(node: SyntaxNode, ctx: ExtractorContext): boolean {
+  const nameNode = getChildByField(node, 'name');
+  if (!nameNode) return true;
+  const rec = ctx.createNode('struct', atomText(nameNode, ctx.source), node, {
+    docstring: getPrecedingDocstring(node, ctx.source),
+    signature: collapseWs(getNodeText(node, ctx.source)).slice(0, 300),
+  });
+  if (rec) {
+    ctx.pushScope(rec.id);
+    for (const field of node.namedChildren) {
+      if (field.type !== 'record_field') continue;
+      const fieldName = getChildByField(field, 'name');
+      if (fieldName) ctx.createNode('field', atomText(fieldName, ctx.source), field);
+    }
+    ctx.popScope();
+  }
+  return true; // field types/defaults are type-position exprs — don't descend
+}
+
+function handleTypeAlias(node: SyntaxNode, ctx: ExtractorContext): boolean {
+  const typeName = getChildByField(node, 'name'); // type_name wrapper
+  const nameNode = typeName ? getChildByField(typeName, 'name') : null;
+  if (nameNode) {
+    ctx.createNode('type_alias', atomText(nameNode, ctx.source), node, {
+      signature: collapseWs(getNodeText(node, ctx.source)).slice(0, 200),
+    });
+  }
+  return true;
+}
+
+function handlePpDefine(node: SyntaxNode, ctx: ExtractorContext): boolean {
+  const lhs = getChildByField(node, 'lhs');
+  const nameNode = lhs ? getChildByField(lhs, 'name') : null;
+  if (nameNode) {
+    ctx.createNode('constant', getNodeText(nameNode, ctx.source), node, {
+      signature: collapseWs(getNodeText(node, ctx.source)).slice(0, 200),
+    });
+  }
+  return true; // the replacement's calls only exist at expansion sites
+}
+
+function handleBehaviour(node: SyntaxNode, ctx: ExtractorContext): boolean {
+  const nameNode = getChildByField(node, 'name');
+  const parentId = ctx.nodeStack[ctx.nodeStack.length - 1];
+  if (nameNode && parentId) {
+    // `-behaviour(x)` implements x's callback contract. Resolves when the
+    // behaviour module is in the repo; OTP behaviours (gen_server, …) simply
+    // stay unresolved.
+    ctx.addUnresolvedReference({
+      fromNodeId: parentId,
+      referenceName: atomText(nameNode, ctx.source),
+      referenceKind: 'implements',
+      line: node.startPosition.row + 1,
+      column: node.startPosition.column,
+    });
+  }
+  return true;
+}
+
+export const erlangExtractor: LanguageExtractor = {
+  functionTypes: ['fun_decl'], // dispatched via visitNode (name lives on the clause)
+  classTypes: [],
+  methodTypes: [],
+  interfaceTypes: [],
+  structTypes: ['record_decl'], // dispatched via visitNode (fields are direct children)
+  enumTypes: [],
+  typeAliasTypes: ['type_alias', 'opaque'], // dispatched via visitNode
+  importTypes: ['import_attribute', 'pp_include', 'pp_include_lib'],
+  callTypes: [
+    'call',
+    'internal_fun', // fun f/1
+    'external_fun', // fun mod:f/1
+    'record_expr', // #rec{...} construction
+    'record_update_expr', // X#rec{...}
+    'record_index_expr', // #rec.field
+    'record_field_expr', // X#rec.field
+  ],
+  variableTypes: [],
+  nameField: 'name',
+  bodyField: 'body',
+  paramsField: 'args',
+
+  // `-module(m)` wraps the file's declarations in a namespace so every
+  // function's qualifiedName is `m::f` — which is exactly the reference shape
+  // the extractCall erlang branch emits for remote calls, so `mod:f(...)`
+  // resolves through matchByQualifiedName with no resolver changes.
+  packageTypes: ['module_attribute'],
+  extractPackage: (node, source) => {
+    const name = getChildByField(node, 'name');
+    return name ? atomText(name, source) : null;
+  },
+
+  extractImport: (node, source) => {
+    if (node.type === 'import_attribute') {
+      const mod = getChildByField(node, 'module');
+      if (!mod) return null;
+      return {
+        moduleName: atomText(mod, source),
+        signature: collapseWs(getNodeText(node, source)).slice(0, 200),
+      };
+    }
+    // pp_include / pp_include_lib — a C-include-style file dependency on a .hrl.
+    const file = getChildByField(node, 'file');
+    if (!file) return null;
+    const headerPath = getNodeText(file, source).replace(/^"/, '').replace(/"$/, '');
+    if (!headerPath) return null;
+    return { moduleName: headerPath, signature: getNodeText(node, source).trim() };
+  },
+
+  visitNode: (node, ctx) => {
+    switch (node.type) {
+      case 'fun_decl':
+        return handleFunDecl(node, ctx);
+      case 'record_decl':
+        return handleRecordDecl(node, ctx);
+      case 'type_alias':
+      case 'opaque':
+        return handleTypeAlias(node, ctx);
+      case 'pp_define':
+        return handlePpDefine(node, ctx);
+      case 'behaviour_attribute':
+        return handleBehaviour(node, ctx);
+      // -spec / -callback: their type expressions parse as `call` nodes;
+      // consume the subtree so the walker doesn't mint bogus call refs.
+      case 'spec':
+      case 'callback':
+        return true;
+      default:
+        return false;
+    }
+  },
+};

src/extraction/languages/index.ts

@@ -31,6 +31,7 @@ import { cfscriptExtractor } from './cfscript';
 import { cfqueryExtractor } from './cfquery';
 import { cobolExtractor } from './cobol';
 import { vbnetExtractor } from './vbnet';
+import { erlangExtractor } from './erlang';
 
 export const EXTRACTORS: Partial<Record<Language, LanguageExtractor>> = {
   typescript: typescriptExtractor,
@@ -59,4 +60,5 @@ export const EXTRACTORS: Partial<Record<Language, LanguageExtractor>> = {
   cfquery: cfqueryExtractor,
   cobol: cobolExtractor,
   vbnet: vbnetExtractor,
+  erlang: erlangExtractor,
 };

src/extraction/tree-sitter-helpers.ts

@@ -84,6 +84,7 @@ function cleanCommentMarkers(comment: string): string {
     .replace(/^\/\/[/!]?\s?/gm, '') // // , and Rust/Swift doc lines /// //!
     .replace(/^--\s?/gm, '') //        Lua/Luau line comments
     .replace(/^#\s?/gm, '') //         Python/Ruby/shell line comments
+    .replace(/^%+\s?/gm, '') //        Erlang line comments (% / %% / %%%)
     .replace(/^\s*\*\s?/gm, '') //     block-comment continuation (* foo)
     .trim();
 }

src/extraction/tree-sitter.ts

@@ -60,6 +60,22 @@ const VUE_STORE_FACTORY_CALLEES = new Set(['defineStore', 'createStore']);
  *  `const actions = {…}` as a store collection — see looksLikeVueStoreFile). */
 const VUE_STORE_FILE_SIGNAL = /\bdefineStore\b|\bcreateStore\b|\bVuex\b|\bmutations\b|\bactions\b|\bgetters\b|\bnamespaced\b/g;
 
+/**
+ * Erlang calls that take their real callee as (Module, Function, Args)
+ * ARGUMENTS — the spawn/apply family. Keys are the callee as the call site
+ * spells it: bare for auto-imported BIFs, `module:function` for remote calls.
+ * Used by the erlang branch of extractCall to lift a static MFA pair into a
+ * call edge (the spawned/applied function is otherwise invisible to the graph).
+ */
+const ERLANG_MFA_CALLS = new Set([
+  'spawn', 'spawn_link', 'spawn_monitor', 'spawn_opt', 'apply',
+  'erlang:spawn', 'erlang:spawn_link', 'erlang:spawn_monitor', 'erlang:spawn_opt', 'erlang:apply',
+  'proc_lib:spawn', 'proc_lib:spawn_link', 'proc_lib:spawn_opt', 'proc_lib:start', 'proc_lib:start_link',
+  'timer:apply_after', 'timer:apply_interval',
+  'rpc:call', 'rpc:cast', 'rpc:async_call',
+  'erpc:call', 'erpc:cast',
+]);
+
 /**
  * Extract the name from a node based on language
  */
@@ -3492,6 +3508,50 @@ export class TreeSitterExtractor {
   /**
    * Extract a function call
    */
+  /**
+   * Whether an Erlang gen_server target expression statically refers to the
+   * module it appears in: `?MODULE`, a macro the file defines as `?MODULE`
+   * (`-define(SERVER, ?MODULE)` — the standard idiom), or the module's own
+   * name as an atom. The self-macro set is memoized per file (single entry —
+   * extraction is file-sequential).
+   */
+  private erlangSelfMacroFile = '';
+  private erlangSelfMacros = new Set<string>();
+
+  private isErlangSelfReference(target: SyntaxNode): boolean {
+    const ownModule = (this.filePath.split('/').pop() ?? '').replace(/\.erl$/, '');
+    if (target.type === 'atom') {
+      return getNodeText(target, this.source) === ownModule;
+    }
+    if (target.type !== 'macro_call_expr') return false;
+    const nameNode = getChildByField(target, 'name');
+    if (!nameNode) return false;
+    const macroName = getNodeText(nameNode, this.source);
+    if (macroName === 'MODULE') return true;
+    if (this.erlangSelfMacroFile !== this.filePath) {
+      this.erlangSelfMacroFile = this.filePath;
+      this.erlangSelfMacros = new Set<string>();
+      let root: SyntaxNode = target;
+      while (root.parent) root = root.parent;
+      for (let i = 0; i < root.namedChildCount; i++) {
+        const form = root.namedChild(i);
+        if (form?.type !== 'pp_define') continue;
+        const lhs = getChildByField(form, 'lhs');
+        const defName = lhs ? getChildByField(lhs, 'name') : null;
+        const replacement = getChildByField(form, 'replacement');
+        if (
+          defName &&
+          replacement?.type === 'macro_call_expr' &&
+          getChildByField(replacement, 'name') &&
+          getNodeText(getChildByField(replacement, 'name')!, this.source) === 'MODULE'
+        ) {
+          this.erlangSelfMacros.add(getNodeText(defName, this.source));
+        }
+      }
+    }
+    return this.erlangSelfMacros.has(macroName);
+  }
+
   private extractCall(node: SyntaxNode): void {
     if (this.nodeStack.length === 0) return;
 
@@ -3543,6 +3603,155 @@ export class TreeSitterExtractor {
       return;
     }
 
+    // Erlang: a local call is `call(expr: atom, args)`; a remote call nests it
+    // under `remote(module: remote_module, fun: call)` — the module qualifier
+    // lives on the PARENT. Remote calls are emitted as `mod::fn`, which is
+    // byte-identical to the qualifiedName the module namespace gives every
+    // function (see packageTypes in languages/erlang.ts), so they resolve via
+    // matchByQualifiedName. A var/macro callee or module (`F(X)`, `?M(X)`,
+    // `Mod:handle(X)`) has no static target — except `?MODULE:fn(X)`, which the
+    // bare name + same-file preference resolves correctly. `fun name/1` /
+    // `fun mod:name/1` values are function REFERENCES (callback registration),
+    // and record construction/update/index/field-access are `references` to the
+    // record's struct node.
+    if (this.language === 'erlang') {
+      const line = node.startPosition.row + 1;
+      const column = node.startPosition.column;
+      const erlAtom = (n: SyntaxNode): string => getNodeText(n, this.source).replace(/^'([\s\S]*)'$/, '$1');
+      if (node.type === 'call') {
+        let callee = getChildByField(node, 'expr');
+        let moduleNode: SyntaxNode | null = null;
+        // remote(module, fun: call) — the shape the grammar produces today; the
+        // node-types also permit call(expr: remote), so handle both nestings.
+        if (node.parent?.type === 'remote') {
+          moduleNode = getChildByField(node.parent, 'module');
+        } else if (callee?.type === 'remote') {
+          moduleNode = getChildByField(callee, 'module');
+          callee = getChildByField(callee, 'fun');
+        }
+        if (callee?.type === 'atom') {
+          const fnBare = erlAtom(callee);
+          let calleeName = fnBare;
+          const moduleExpr = moduleNode ? getChildByField(moduleNode, 'module') : null;
+          if (moduleExpr?.type === 'atom') {
+            calleeName = `${erlAtom(moduleExpr)}::${calleeName}`;
+          } else if (moduleExpr) {
+            // Non-atom module qualifier. `?MODULE:f(X)` targets THIS module —
+            // keep the bare name so same-file preference resolves it. Anything
+            // else (`Mod:f(X)`) is behaviour-style dynamic dispatch with no
+            // static target: emitting the bare name would link an arbitrary
+            // same-named function, so stay silent instead.
+            const macroName =
+              moduleExpr.type === 'macro_call_expr' ? getChildByField(moduleExpr, 'name') : null;
+            if (!macroName || getNodeText(macroName, this.source) !== 'MODULE') return;
+          }
+          this.unresolvedReferences.push({
+            fromNodeId: callerId,
+            referenceName: calleeName,
+            referenceKind: 'calls',
+            line,
+            column,
+          });
+          // gen_server self-dispatch: `gen_server:call(?SERVER, Msg)` /
+          // `gen_server:cast(?MODULE, Msg)` — the OTP API-wrapper idiom (a
+          // module's public functions wrap gen_server requests to itself, and
+          // the real work happens in its own handle_call/handle_cast). The
+          // target is static when the first argument is ?MODULE, a macro the
+          // file defines as ?MODULE (the standard `-define(SERVER, ?MODULE)`),
+          // or the module's own name as an atom — emit the qualified callback
+          // ref so the module's public API connects to its handlers. Any other
+          // target (pid/var/registered name of another process) stays silent.
+          if (
+            moduleExpr?.type === 'atom' &&
+            erlAtom(moduleExpr) === 'gen_server' &&
+            (fnBare === 'call' || fnBare === 'cast' || fnBare === 'send_request')
+          ) {
+            const argsNode = getChildByField(node, 'args');
+            const target = argsNode?.namedChild(0) ?? null;
+            if (target && this.isErlangSelfReference(target)) {
+              const ownModule = (this.filePath.split('/').pop() ?? '').replace(/\.erl$/, '');
+              if (ownModule) {
+                this.unresolvedReferences.push({
+                  fromNodeId: callerId,
+                  referenceName: `${ownModule}::${fnBare === 'cast' ? 'handle_cast' : 'handle_call'}`,
+                  referenceKind: 'calls',
+                  line,
+                  column,
+                });
+              }
+            }
+          }
+          // MFA-in-argument dispatch: the spawn/apply family names its real
+          // callee in ARGUMENT position — `proc_lib:spawn_link(?MODULE,
+          // request_process, [Req, Env, Middlewares])` — so the walker above
+          // sees only the spawn itself and the spawned function ends up with
+          // zero callers (measured on cowboy: request_process had no incoming
+          // edges and the agent Read the file to find it). When the (Module,
+          // Function) pair is static, lift it as a call edge. The pair is
+          // found positionally-agnostically (first adjacent module-atom/
+          // ?MODULE + atom pair) so every arity variant works: spawn/3,
+          // spawn(Node,M,F,A)/4, timer:apply_after(Time,M,F,A),
+          // rpc:call(Node,M,F,A). A var module or fun stays silent.
+          const familyKey = moduleExpr?.type === 'atom' ? `${erlAtom(moduleExpr)}:${fnBare}` : fnBare;
+          if (ERLANG_MFA_CALLS.has(familyKey)) {
+            const argsNode = getChildByField(node, 'args');
+            const argExprs = argsNode ? argsNode.namedChildren : [];
+            for (let i = 0; i + 1 < argExprs.length; i++) {
+              const m = argExprs[i]!;
+              const f = argExprs[i + 1]!;
+              if (f.type !== 'atom') continue;
+              const isLocalModule =
+                m.type === 'macro_call_expr' &&
+                getChildByField(m, 'name') !== null &&
+                getNodeText(getChildByField(m, 'name')!, this.source) === 'MODULE';
+              if (m.type !== 'atom' && !isLocalModule) continue;
+              this.unresolvedReferences.push({
+                fromNodeId: callerId,
+                referenceName: isLocalModule ? erlAtom(f) : `${erlAtom(m)}::${erlAtom(f)}`,
+                referenceKind: 'calls',
+                line: f.startPosition.row + 1,
+                column: f.startPosition.column,
+              });
+              break;
+            }
+          }
+        }
+        return;
+      }
+      if (node.type === 'internal_fun' || node.type === 'external_fun') {
+        const funNode = getChildByField(node, 'fun');
+        if (funNode?.type !== 'atom') return; // fun Mod:F/A with var parts — dynamic
+        let refName = erlAtom(funNode);
+        if (node.type === 'external_fun') {
+          const moduleWrapper = getChildByField(node, 'module');
+          const moduleAtom = moduleWrapper ? getChildByField(moduleWrapper, 'name') : null;
+          if (moduleAtom?.type !== 'atom') return;
+          refName = `${erlAtom(moduleAtom)}::${refName}`;
+        }
+        this.unresolvedReferences.push({
+          fromNodeId: callerId,
+          referenceName: refName,
+          referenceKind: 'references',
+          line,
+          column,
+        });
+        return;
+      }
+      // record_expr / record_update_expr / record_index_expr / record_field_expr
+      const recordName = getChildByField(node, 'name');
+      const recordAtom = recordName?.type === 'record_name' ? getChildByField(recordName, 'name') : null;
+      if (recordAtom?.type === 'atom') {
+        this.unresolvedReferences.push({
+          fromNodeId: callerId,
+          referenceName: erlAtom(recordAtom),
+          referenceKind: 'references',
+          line,
+          column,
+        });
+      }
+      return;
+    }
+
     // Ruby `call` nodes use `receiver` + `method` fields (tree-sitter-ruby), not
     // the `object`/`name`/`function` fields the branches below expect — so
     // without this they fell through to the generic path, which took the