fix(archtest): resolve the recursion pin's calls with go/types

[OmarAlJarrah]

Summary

internal/archtest's recursion pin (TestLoweringRecursion_IsOnlyTheKnownCycles) built the lowering call graph from go/parser output alone, so calleesOf, declOf and typeNameOf reconstructed by hand what each name refers to. That resolver had been extended twice for missed edges (#210, #323) and still missed one class: a method called on a value bound by a multi-result assignment (probe, merges, err := …; probe.fillFrom(…)). The edge probeFromMapping -> sniffProbe.fillFrom is absent from the graph on main, and a new cycle written in that shape would never be pinned.

This loads the four lowering packages with golang.org/x/tools/go/packages and takes every edge from types.Info.Uses, which is what #797 proposed. The hand-written resolver and its TestCalleesOf_* / TestDeclOf_* / TestTypeNameOf_* tests are removed (git diff --stat origin/main -- internal/archtest/recursion_test.go shows how much shorter recursion_test.go gets). One table of planted sources, TestLoweringCallGraph_FollowsWhatTheTypeCheckerResolves, covers the shapes a parser-only resolver gets wrong or can only guess at: multi-result binds, function values and method values, shadowing locals, closure bodies, methods on instantiated generic types, receivers spelled through an alias, promoted methods, range-variable and type-switch bindings, and interface calls (which still record nothing). It also holds what must stay out of the graph: a field, or a composite-literal key, that shares a node's name.

Result on the current tree

Both graphs were dumped edge by edge, on main and on this branch:

• The node set is identical, and no edge was removed.
• 39 edges were added: the fillFrom edge above, plus 38 calls between the packages read (run -> LowerService, fillParamType -> CarriedRef, …). The old graph skipped cross-package calls on purpose, since none can close a cycle: Go refuses the import that would let the callee's package call back. They come along with type resolution but are not a fix, and the doc on loweringCallGraph keeps that argument.
• All seven loweringRecursions sets are reproduced unchanged.

go/callgraph/static and cha were tried too and rejected. static loses function-value edges (the schema set shrinks from 39 members to 10). cha over-approximates through interfaces and invents a new set. Uses matches the semantics the pin already had.

Cost

• Warm cache: the whole internal/archtest package goes from about 0.49 s to 0.71 s.
• Cold cache, in gate order (go build ./... first): 0.46 s to 0.89 s, about 1 s more wall time including building x/tools into the test binary.
• Dependency: golang.org/x/tools v0.52.0 becomes a direct requirement. Only internal/archtest's tests import it, and go list -deps ./... shows no production package depending on it, so no rules entry changes (rules audits production files only, and internal/archtest has none). x/tools requires x/sync v0.24.0. The bump from v0.23.0 differs only in errgroup_test.go, so production compiles the same errgroup source. x/mod v0.42.0 is new. x/net and x/sys move to v0.61.0 and v0.49.0 because x/tools requires them; go list -deps -test ./... shows no package in this module compiles either.
• Toolchain coupling: x/tools reads the toolchain's export data, and v0.49.0 cannot read Go 1.27.2's ("export data version 5 is greater than maximum supported version 4"). A future toolchain bump that outruns it fails this test loudly through p.Errors. The Makefile comment above GO_VERSION now lists x/tools among the things to bump together.

Closes #797

Test plan

• go test ./internal/archtest and go test -race ./internal/archtest pass, and TestDocComments_StayUnderTheCap is green.
• go build ./..., go vet ./... and go mod tidy -diff are clean, gofmt -l is empty, and golangci-lint v2.14.0 (the Makefile pin) reports 0 issues over ./... on a fresh cache.
• Mutations, each built before running, with a no-op control that survived:
  • Dropping either types.Unalias in nodeName reddens the aliased-receiver case.
  • Dropping .Origin() in calleesIn reddens the generic-method case.
  • Keying methods by bare name reddens every graph test.
  • Skipping exported callees reddens TestLoweringRecursion_IsOnlyTheKnownCycles.
  • Skipping closure bodies, counting a field or any variable as an edge, recording calls only, dropping key-value pairs, refusing promoted methods, and refusing range-variable and type-switch bindings each redden the planted table on its own (run with -run on that test), and a no-op rewrite of calleesIn survives it.
• Edge-by-edge diff of the old and new graphs on the same tree, with the numbers given above.
• Full gate (coverage, fuzz, bench smoke) is left to CI.