tests: remove the DNS stand-ins from the watcher and resolver tests (closes #159)

The watcher tests used a stand-in resolver and the resolver timeout test
a stand-in DNS client, against the rule that DNS is never mocked. The
watcher tests now run the real resolver against live DNS. A change is
tested by saving values live DNS never returns (names under .invalid,
192.0.2.1) in the state a check starts from, or by marking a real
nameserver failed. The timeout test queries 192.0.2.1, where nothing
answers. The live-DNS retry and concurrency limit moved from the
resolver tests to internal/livedns, so both packages share them.
NewFromLoggerWithClient had no other use and is gone. TESTING.md and
the DNSClient comment now state the README's rule.

Model: opus-5-5

internal/livedns/livedns.go

@@ -0,0 +1,119 @@
+// Package livedns runs the live DNS operations of tests. Every test in
+// this project that needs DNS resolves against the real, live DNS —
+// see TESTING.md. Nothing here mocks, fakes, stubs, records or replays
+// DNS, and nothing here skips a test: it only changes *how* the live
+// queries are issued, so that a single dropped UDP packet or one slow
+// authoritative server does not turn correct code into a red build.
+//
+// Two mechanisms:
+//
+//  1. Bounded concurrency. Tests run in parallel and the build hosts
+//     have many cores, so without a limit every test starts its own
+//     iterative resolution at the same instant and they all hit the
+//     first root server within a few milliseconds of each other. Root
+//     servers rate-limit that, which shows up as a different arbitrary
+//     subset of tests failing on each run. Run caps how many live
+//     operations are in flight at once in one test binary.
+//
+//  2. Retry with exponential backoff. Each live operation gets several
+//     attempts with its own timeout. The retry condition is strictly
+//     transport-level — "did a nameserver answer at all" — never the
+//     assertion the test is making. Code that answers incorrectly
+//     still fails on the first attempt.
+package livedns
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+)
+
+const (
+	// attempts is how many times a live DNS operation is attempted
+	// before the test fails.
+	attempts = 3
+
+	// AttemptTimeout bounds one attempt. Worst case for an operation
+	// is attempts * AttemptTimeout plus the backoff — about 26
+	// seconds, well inside the 90-second `go test -timeout` backstop
+	// even when several operations exhaust their attempts.
+	AttemptTimeout = 8 * time.Second
+
+	// backoffBase is the delay after the first failed attempt; it is
+	// multiplied by backoffFactor each time.
+	backoffBase = 500 * time.Millisecond
+
+	// backoffFactor is the exponential backoff multiplier.
+	backoffFactor = 2
+
+	// Concurrency caps how many live operations may be in flight
+	// across one test binary at once.
+	Concurrency = 6
+)
+
+// gate bounds concurrent live operations. It has to be package scoped:
+// the whole point is that it is shared by every parallel test in the
+// test binary.
+//
+//nolint:gochecknoglobals // package-wide live query rate limit
+var gate = make(chan struct{}, Concurrency)
+
+// ErrNoAnswer reports that a live operation produced no usable answer,
+// which is retried rather than asserted on.
+var ErrNoAnswer = errors.New("no answer from live DNS")
+
+// Run executes one attempt of a live operation, holding a slot in gate
+// for its duration and bounding it with its own timeout.
+func Run(op func(ctx context.Context) error) error {
+	gate <- struct{}{}
+	defer func() { <-gate }()
+
+	ctx, cancel := context.WithTimeout(
+		context.Background(), AttemptTimeout,
+	)
+	defer cancel()
+
+	return op(ctx)
+}
+
+// Retry runs op until it reports success, retrying failures with
+// exponential backoff, and fails the test if every attempt fails. op
+// returns an error only for a failure to obtain an answer — never for
+// an answer the test disagrees with, which belongs in an assertion so
+// that it fails immediately. op stores whatever it obtained where its
+// caller can find it.
+func Retry(
+	t *testing.T,
+	what string,
+	op func(ctx context.Context) error,
+) {
+	t.Helper()
+
+	var last error
+
+	backoff := backoffBase
+
+	for attempt := range attempts {
+		if attempt > 0 {
+			t.Logf(
+				"%s: attempt %d of %d failed (%v), "+
+					"retrying in %s",
+				what, attempt, attempts, last, backoff,
+			)
+			time.Sleep(backoff)
+
+			backoff *= backoffFactor
+		}
+
+		last = Run(op)
+		if last == nil {
+			return
+		}
+	}
+
+	t.Fatalf(
+		"%s: no answer after %d live attempts: %v",
+		what, attempts, last,
+	)
+}

internal/livedns/livedns_test.go

@@ -0,0 +1,103 @@
+package livedns_test
+
+import (
+	"context"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+
+	"sneak.berlin/go/dnswatcher/internal/livedns"
+)
+
+// Tests for the retry and the concurrency limit themselves. They
+// perform no DNS resolution of any kind.
+
+func TestRetryRecoversFromTransientFailure(t *testing.T) {
+	t.Parallel()
+
+	const wantAttempts = 2
+
+	attempts := 0
+
+	livedns.Retry(t, "transient", func(_ context.Context) error {
+		attempts++
+
+		if attempts < wantAttempts {
+			return livedns.ErrNoAnswer
+		}
+
+		return nil
+	})
+
+	assert.Equal(t, wantAttempts, attempts)
+}
+
+func TestRetryGivesEachAttemptADeadline(t *testing.T) {
+	t.Parallel()
+
+	livedns.Retry(t, "deadline", func(ctx context.Context) error {
+		deadline, ok := ctx.Deadline()
+		assert.True(t, ok, "attempt should carry a deadline")
+
+		remaining := time.Until(deadline)
+
+		assert.LessOrEqual(t, remaining, livedns.AttemptTimeout)
+
+		// Lower bound too: without one this passes for a
+		// deadline far shorter than intended, which would
+		// silently turn every live attempt into an instant
+		// timeout.
+		assert.Greater(t, remaining, livedns.AttemptTimeout/2)
+
+		return nil
+	})
+}
+
+func TestRunBoundsConcurrency(t *testing.T) {
+	t.Parallel()
+
+	const workers = 24
+
+	var (
+		mu       sync.Mutex
+		wg       sync.WaitGroup
+		inFlight int
+		maxSeen  int
+	)
+
+	wg.Add(workers)
+
+	for range workers {
+		go func() {
+			defer wg.Done()
+
+			_ = livedns.Run(func(_ context.Context) error {
+				mu.Lock()
+				inFlight++
+
+				if inFlight > maxSeen {
+					maxSeen = inFlight
+				}
+				mu.Unlock()
+
+				time.Sleep(time.Millisecond)
+
+				mu.Lock()
+				inFlight--
+				mu.Unlock()
+
+				return nil
+			})
+		}()
+	}
+
+	wg.Wait()
+
+	assert.Positive(t, maxSeen)
+	assert.LessOrEqual(
+		t, maxSeen, livedns.Concurrency,
+		"live queries must stay under the package-wide gate",
+	)
+}