Fix a race in TestClient

[jdtw]

What happened

CI failed on the merge of #30. That PR added only four files under frontend/ — no Go, no go.mod, no workflow changes — so it didn't cause this. TestClient has been racy all along and happened to lose on that run.

client_test.go:50: Get(foo) returned Get "http://localhost:46697/api/links/foo":
    dial tcp [::1]:46697: connect: connection refused; want err not found

The race

addr := fmt.Sprintf("localhost:%d", getFreePort(t))   // binds, then CLOSES
s := &http.Server{Addr: addr, ...}
go func() { s.ListenAndServe() }()                     // binds asynchronously
c := New("http://"+addr, signer)
c.Get("foo")                                           // dials immediately

Nothing orders the bind before the first request. On an idle machine the goroutine wins; on a loaded CI runner it doesn't. There's a second, narrower race too: closing the listener before rebinding leaves a window for another process to take the port.

The [::1] in the error is a symptom of the same design — localhost was reassembled from a port number rather than using the address actually bound, so the client could resolve to a different stack than the server used.

The fix

httptest.NewServer binds before it returns and hands back the address it actually bound. That closes both races and the IPv4/IPv6 mismatch at once, and drops getFreePort, the sync.WaitGroup, and the manual Shutdown. pkg/frontend/server_test.go already tests this way, so this makes the two consistent.

Verification

I could not reproduce the failure locally — expected for a scheduling race on an idle machine. I tried -count=500, GOMAXPROCS=1, and saturating every core; it passed throughout, which is why the fix is reasoned from the code rather than from a red test.

After the change: 500 runs, 50 runs under -race, 300 under GOMAXPROCS=1, the full unit suite, and the end-to-end ./test.sh all pass.

🤖 Generated with Claude Code