test+refactor: real sidecar test, deterministic test sleeps, slog step line (#22)

* test(sidecar): replace assertTrue(true) placeholder with real server test

MainTest.mainExists() always passed and inflated the green-check count.
DriverServiceTest covers RPCs, but SidecarServer start/stop had no
coverage. Drop the placeholder and add SidecarServerTest that binds to
port 0, asserts a real ephemeral port, and stops cleanly.

* test(agent): drop 50ms sleep before cancel in TestServer_AcceptCancelsOnContext

Accept's closeListenerOnCancel watcher closes the listener as soon as
ctx fires, regardless of whether the outer Accept has reached
listener.Accept() yet. The sleep was a CI-flake surface (50ms is not
enough on a slow runner), and dropping it still exercises the same
outcome — Accept returns with ctx.Err() after cancellation.

Stable across 50x -count runs.

* test(agent): replace 2s sleep with done-chan in TestConn_SnapshotTimesOutIfSDKSilent

The silent-SDK fake held the connection open via time.Sleep(2s), which
coupled the test's wall clock to the server's 200ms snapshot-timeout
assertion. Swap for a done channel closed by t.Cleanup — the goroutine
exits when the test ends, independent of timing.

* test(maestro): make WaitForHealth_PollsUntilReady deterministic

Replace the 50ms wall-clock sleep that flipped healthReady with a
healthReadyAfterCall counter in the fake server. The handler returns
ready=true once healthCalls reaches the threshold, so the test's
"at least 2 polls before ready" assertion is satisfied by call
count rather than a race between the flip goroutine and the 25ms
poll loop.

* refactor(runner): route per-step progress through slog instead of fmt.Printf

The runner already carries a *slog.Logger for warnings (logger.Warn on
decode failures, predicate errors). The per-step status line was the
outlier — a bare fmt.Printf that wrote to os.Stdout unconditionally,
bypassing both the injected logger and any caller-configured writer.

Switch it to logger.Info("step", "index", ..., "screen", ..., "nodes", ...).
The caller (cmd/uatu) is responsible for wiring a logger whose handler
renders to the right stream; the next commit adds that wiring.

* feat(cli): render runner progress via a thin slog handler on stdout

progressHandler writes Info records as "msg key=value ..." and prefixes
warnings/errors with their level, matching the prose style of the
surrounding CLI status prints. Wired into the runner via
runner.Options.Logger so the per-step status line still lands on stdout
without slog's default time= / level= framing.

internal/runner/runner.go

@@ -104,8 +104,7 @@ func Run(ctx context.Context, options Options) (Summary, error) {
 		if screenErr != nil {
 			logger.Warn("screen snapshot decode failed", "step", stepIndex, "err", screenErr)
 		}
-		fmt.Printf("step %d: screen=%q hierarchy=%d nodes\n",
-			stepIndex, screen, treeSize)
+		logger.Info("step", "index", stepIndex, "screen", screen, "nodes", treeSize)
 		verdicts := options.Verifier.EvaluateProperties()
 		violations := violationNames(verdicts)
 		for _, name := range violations {