fix(go): a call through a type assertion reaches the asserted type's method

[colbymchenry]

Problem

A Go call through a type assertion, like srv.(KVServer).Range(ctx, in), reaches the resolver by its bare member name (Range), recorded at the column where its receiver expression starts, the way every call through an expression is. Resolution matched the name alone:

• every gRPC handler protoc-gen-go-grpc emits went to the Unimplemented…Server stub beside the server interface (etcd's api/etcdserverpb/rpc_grpc.pb.go:223 → UnimplementedKVServer::Range, exact-match 0.7; 72 such edges in etcd, 3 in kratos, 95 in grpc-go);
• other assertion calls went to a namesake (cfg.ServerFeatureGate.(featuregate.MutableFeatureGate).Set(…) → v2store's Store::Set; stream.(grpc.ClientStream).SendMsg(…) → an example's wrappedStream) or to nothing (p.(Pausable).Pause(), srv.(LockServer).Lock(…), c.Reader.(*pipe).Close()).

Fix

matchGoAssertedCall (name-matcher) runs early in resolveOneInner, ahead of the framework, import and name strategies and of the built-in filter (which dropped a method named close or copy). It reads the call's chain back from its column: the operand, then each selector, assertion, call and index, past strings, runes and comments, and onto the next line after a trailing ., where Go inserts no semicolon. When the link of the call's name follows an assertion, the call is a method of the asserted type, found where Go finds it:

• a bare T or *T (type arguments dropped) in the call's own package, or in a project package the file dot-imports;
• a pkg.T in the imported project package;
• then resolveMethodOnType with that package directory, the package-scoped lookup fix(go): resolve unexported receivers and nested modules within the right package (#2323, #2322) #2361 added: the type's own method, the one its interface declares, or one promoted from a type it embeds; an alias is followed to the type it names (fix(go): a type alias type A = B is indexed and links to the type it names #2417).

A type found in none of those (outside the project, like http.Flusher, or predeclared, like error), an alias of an outside type (type Ctx = context.Context) and a type literal (interface{ Flush() }) link nothing. Two calls of one name in one chain (b.(*Builder).Add(1).Add(2)) share the column and can't be told apart; both are taken for the one made through the assertion, whose edge is real either way.

Kernel: no change needed. codegraph-kernel/src/go.rs extract_call falls to its _ arm for a type_assertion_expression receiver exactly as the TS extractor's final else does, so both emit the bare name at the call's start. The new test pins that on both backends (LF and CRLF), and the two paths' indexes agree on all 298 assertion sites in etcd and grpc-go (ref names, columns and statuses, edge targets).

Validation

Regression test: __tests__/go-type-assertion-calls.test.ts, 18 tests. The extraction check runs on the kernel and wasm; the indexed module runs on the default path and on wasm. Namesakes sit where name matching looked first: a stub beside each interface, a same-file type with each method name, and an alpha package that sorts ahead of storage. With main's resolver, 12 of the 14 resolution tests fail (6 per backend); the extraction checks and the guard that a later link of the chain is left alone pass there by design. Each guarded piece was mutation-checked: dropping dot-import lookup, string skipping, the hook's place ahead of the built-in filter, the null for a type the package doesn't declare, the alias check, or leaving two same-name links of a chain to name matching each turns a test red.

Before/after edge diffs (site-keyed on (source, line, column, ref name), main ed199e6 vs this branch, same kernel; every changed site triaged against the source: the type asserted at the call, the package Go finds it in from go.mod and the file's imports, and each side's target):

repo	changed call sites	retargeted	added	dropped	same target, metadata only
etcd	117	99	9	1	8
grpc-go	158	123	2	1	32
harbor	19	0	0	7	12
kratos	6	3	0	0	3

• Retargeted (225): every one now lands on a method of the asserted type in the right package. etcd: 72 off Unimplemented…Server stubs, 17 off AuthClient (a client interface sharing the RPC names), 10 MutableFeatureGate.Set calls off v2store's Store::Set and cache's ready::Set. grpc-go: 95 off stubs, 28 off other namesakes (balancer.ConfigParser.ParseConfig off each builder's own, grpc.ServerStream.RecvMsg off an example's wrappedStream, *testutils.TestSubConn off a wrapper in the caller's package, …). kratos: the 3 GreeterServer handler calls.
• Added (11), all correct: etcd p.(Pausable).Pause/Resume, srv.(LockServer).Lock/Unlock ×4, iv.Val.(watcherSet).add (a defined map type's method), st1.(EtcdState).Equal, st1.(nonDeterministicState).Equal; grpc-go c.Reader.(*pipe).Close() (it used to collapse onto the calling conn.Close) and w.s.Handler.(*grpc.Server).Stop() (the root package's Server).
• Dropped (9), all outside the project: etcd w.(http.Flusher).Flush() (was a test fake's Flush), grpc-go (*v3orcapb.OrcaLoadReport).GetRequestCost() (cncf/xds; was interop's TestOrcaReport), harbor 7 × suite.userManager.(*mockUser.Manager).On(…): mockery mocks that embed testify's mock.Mock, so On is testify's (the old edges went to harbor's private mockable interface by name).
• Metadata only (55): same target, now instance-method 0.9 instead of exact-match; 5 of harbor's resolve through the *cached.BaseManager a cached manager embeds, as before.
• The changed sites were the same, site for site, on the previous main (d8a7f86), before fix(go): a call through an interface reaches a method the struct gets by embedding #2414, fix(frameworks): a framework resolves only references written in its own languages #2415, fix(go): a type in a signature or composite literal links a type, never a method of its name #2416, fix(go): a type alias type A = B is indexed and links to the type it names #2417 and fix(go): a defined type implements the interfaces its methods satisfy #2419 changed Go resolution.

Flows stay connected. On main, the stub-bound handler calls reached implementations through the name-only go-grpc-stub-impl bridge; now they land on the interface method and ride interface-impl dispatch (_KV_Range_Handler → KVServer.Range → kvServer.Range in v3rpc/key.go in codegraph_explore's Flow). For every moved site I compared the implementations its target dispatches to in one hop: those reached before but not after are 415 on etcd and 18 on grpc-go, and none is an implementer of the interface the call lands on. They are clients (retryKVClient, Election), server-to-client adapters (kvs2kvc, es2ec), printers, mocks and EtcdServer (still reached one hop later, through kvServer); none embeds an Unimplemented…/Unsafe…Server or has the interface's method set.

Cost: with each file's lines loaded (the resolver's other Go checks read the same lines), the check costs about what reading the ref's line does: 5–70 ms per pass over every Go call ref of these repos (13k–97k refs).

Suite: the box ran at 100% CPU with 100–220 other node processes throughout. In parallel, 104 of 6,684 tests failed (100 timeouts, 2 EBUSY/EEXIST teardowns, a writer-lock state check and a parse-time bound). A rerun of those 40 files with 3 workers and long timeouts left 3 (daemon-pid-reuse EBUSY, function-ref's Python-only #1820 case against its own 60 s limit, mcp-writer-lock), and those 3 files pass alone. tsc is clean.

Not in this PR

• A local bound from an assertion, v, ok := x.(T) then v.M(), still has no inferred type: Go's receiver patterns read := composite literals, var and parameters, not assertions. About 90 such locals across these four repos (49 in grpc-go).
• A defined type over an interface (type MyServer pb.KVServer) has that interface's methods; an assertion to it finds no method of its own and links nothing. Not seen in these repos.

🤖 Generated with Claude Code