[colbymchenry]

Last follow-up from #2678, found while doing #2680. Variables declared in a Go var ( … ) block were not indexed at all. tree-sitter-go wraps a var group's specs in a var_spec_list node, unlike a const ( … ) group's, which sit directly in the declaration. Both extractors read only the specs directly inside a declaration, so etcd, prometheus and gin had no node for any grouped variable. What followed from that:

• the calls and composite literals in their values were never recorded, so a struct used only there looked unused (gin's bsonBinding in var ( BSON BindingBody = bsonBinding{} … ));
• code reading them had nothing to reference;
• a call through one, like prometheus's fPointPool.Get() or etcd's slowApplies.Inc(), was matched by its method name alone. fix(go): a call on a package-level variable or constant links through the type it is declared with #2667 reads a package-level variable's declared type, but only for variables that have a node.

Change

• In the Go branch of variable extraction (extractVariable in TS, extract_variable in the kernel), the specs inside a var_spec_list are read like a declaration's own. Each member becomes a variable the file contains, with its initializer walked under it, as a variable declared on its own already is.
• Docs follow fix(go): a constant in a const ( … ) group keeps its own doc comment #2680: a member's own doc comment, or the group's when it has none.
• Unchanged, as for ungrouped declarations: a spec naming several variables (a, b = 1, 2) still makes a node for the first name only. That limit is shared with const and ungrouped var and is not touched here.

Verification

• New __tests__/go-var-groups.test.ts, kernel and wasm, each on LF and CRLF:
  • fails 4/4 on main 6e68a11 (no node for a grouped variable) and passes 4/4 here;
  • covers each member's node and contains edge, own or group doc, a call in an initializer recorded under its variable, and same-file reads becoming references.
• The torture.go parity fixture gains a var group with a call and a function value in its values; kernel-tsjs-parity passes 99/99. The Go doc tests from fix(go): a type declared on its own keeps its doc comment #2438, fix(go): a comment a blank line separates from a declaration is not its doc #2678, fix(go): a comment directive like //go:embed is not part of a declaration's doc #2679 and fix(go): a constant in a const ( … ) group keeps its own doc comment #2680 pass.
• Full suite on the branch synced with main 16d1c95, with the kernel built and CODEGRAPH_KERNEL_EXPECT=1 (--maxWorkers=4 --minWorkers=1): 7,523 passed, 0 failed, 80 skipped (7,603 tests, 0 failed files).

A/B on real repos

main 6e68a11 vs this branch. #2681, merged in by the sync, changes only TypeScript-family file discovery. Repos: etcd 6782f522, prometheus 0b57789d, gin 0f09c3a9.

Kernel and wasm indexes are byte-identical on both arms (nodes, edges, unresolved refs, files, docstrings). No node is removed; every new node is a grouped variable.

	etcd	prometheus	gin
new variable nodes (+ contains edges)	1,080	334	78
same-file value references to them	1,105	468	54
edges from their values (calls, instantiates, value/function refs)	612	214	62
calls now landing on a function-valued variable	134	0	0
calls resolved for the first time	7	0	1
calls re-resolved	0	8	0
calls no longer linked	38	19	0
function refs no longer linked	1	0	0

How each row was checked:

• Value references (1,276 / 527 / 55, those from new nodes included): every one is same-file, and the reading function's body names the variable.

• Calls landing on variables are calls of a function-valued variable, e.g. grpc-gateway's generated forward_KV_Range_0(…) (= runtime.ForwardResponseMessage) and etcd tests' clusterTestCases(). Main already links calls to ungrouped function-valued variables the same way (9 in etcd).

• Resolved for the first time:

  • linuxTryLockFile(…) and openWALFilesFn(…) reach the function their variable is set to (= flockTryLockFile, = openWALFiles);
  • Meta.Name() reaches backend.Bucket.Name;
  • s.Write(…) with var ( … s mvcc.KV ) reaches mvcc.KV.Write;
  • gin's MsgPack.Bind(…) reaches Binding.Bind.

• Re-resolved, all from wrong to right:

  • prometheus fPointPool.Get()/.Put() (×6): a zeropool.New(…) value, from chunkenc.Pool to zeropool.Pool;
  • testHistogram.Copy() (×2): from FloatHistogram.Copy to Histogram.Copy.

• No longer linked, all wrong before (each a call through a grouped variable whose value or type comes from outside the project, which fix(go): a call on a package-level variable or constant links through the type it is declared with #2667 now reads):

  • etcd: 33 Prometheus counter/gauge .Inc() calls linked to v2store's Stats.Inc; a sync.Once's .Do linked to the client KV.Do; 3 cobra/text/template .Execute linked to migrationPlan.Execute; a regexp's .Match linked to Prefix.Match;
  • prometheus: 9 regexp .MatchString linked to FastRegexMatcher.MatchString; 9 Prometheus counter .Inc()/.Dec() linked to noopMetric; fgprof.Handler()'s .ServeHTTP linked to readHandler.ServeHTTP.

• The function ref tt.rangeLimit in an etcd test, a struct field, had been linked to the txn package's rangeLimit function by name. A grouped rangeLimit variable in tools/benchmark now shares that name, so the name-only guess no longer picks one.

• Edges from the new variables' values follow the resolver's existing rules. Composite literals instantiate their types, and calls reach boolPtr, getGoGC, MustNewRegexp, the prompb converters and others. A few inherit existing name guesses, which main already makes for the same calls in ordinary functions of the same files:

  • prometheus testParser.ParseExpr(…) goes to a test type's ParseExpr, as main's three uses in rules/recording_test.go do;
  • a kingpin a.Flag(…).String() goes to rwProtoMsgFlagParser.String, as 18 of main's do;
  • etcd's txn.End() on a kv.Read(…) result goes to lease's TxnDelete.End, as main's three uses in kv_test.go do.

  Open PR fix(go): type receivers from package-qualified parameters and call results; no guess for outside types #1954 is on call-result typing, the gap behind these.

• Dead code: the report (default kinds, and with exported symbols) is identical. Underneath, 21 / 3 / 11 symbols that had no incoming edge now have one from a grouped variable's value: gin's bsonBinding, formBinding, headerBinding and the other binding structs, timeCodec; etcd's NewIntegrationRunner/NewUnitRunner, GlobalFlags; prometheus's getGoGC. So impact on them now reaches the variables and their readers. The report's own text check had kept them off it.

• Kernel parity (kernel-parity.mjs --lang go): 0 diffs on all three repos. etcd 1065/1105, the same 40 erroring files deferred as on main; prometheus 746/746; gin 99/99.

• Cost: indexing time unchanged within noise (etcd 10→11 s, prometheus 12→14 s, gin 2→3 s on this box).

🤖 Generated with Claude Code