feat(stream): benchmark software encode ceiling at startup

Replace the guessed transcode ceiling (CPU->1080, GPU->2160) with a measured
one. HW encoders still return 2160 instantly. A software-only host runs a
bounded encode benchmark — 3s testsrc2 through the real libx264 superfast
settings at 1080/720/480, top-down — and reports the rung it sustains at
>=1.5x realtime (margin for real decode + busier content).

Fixes risk 2: a weak NAS/old CPU that is ffmpeg-capable but can't keep up
with a 1080p software encode no longer advertises a 1080 ceiling, so
decideStreamPlan routes oversized sources to an external player instead of a
stuttering transcode. Floors at 480; each probe is timeout-bounded so a
wedged ffmpeg can't stall daemon startup.

internal/cmd/daemon.go

@@ -156,10 +156,15 @@ func runDaemonStart() error {
 	hwDiag := engine.DetectHWAccelDiagnostic(probeCtx, ffmpegResolved)
 	log.Println(hwDiag.LogLine())
 	hwAccelPick := hwDiag.Pick
-	maxTranscodeHeight := 1080
-	if hwAccelPick != engine.HWAccelNone {
-		maxTranscodeHeight = 2160
-	}
+	// Measure the real transcode ceiling instead of guessing from the backend.
+	// HW encoders return 2160 instantly; a software-only host runs a bounded
+	// encode benchmark so a weak NAS/CPU reports the rung it can actually
+	// sustain (720/480) and the web side routes oversized sources to an
+	// external player instead of a stuttering transcode. Own timeout — the 10 s
+	// probeCtx above is sized for the quick diagnostic, not three encode rungs.
+	benchCtx, benchCancel := context.WithTimeout(context.Background(), 45*time.Second)
+	maxTranscodeHeight := engine.BenchmarkMaxTranscodeHeight(benchCtx, ffmpegResolved, hwAccelPick)
+	benchCancel()
 
 	// Create daemon config
 	daemonCfg := agent.DaemonConfig{

internal/engine/encode_benchmark.go

@@ -0,0 +1,106 @@
+package engine
+
+import (
+	"context"
+	"log"
+	"os/exec"
+	"strconv"
+	"time"
+)
+
+// benchmarkRung is a candidate transcode-height ceiling plus the 16:9 frame
+// size used to measure whether a software encoder sustains it.
+type benchmarkRung struct {
+	height int
+	width  int
+}
+
+// softwareBenchmarkRungs are tested high→low. The frame sizes match the real
+// streaming output tiers; the H.264 level / macroblock math in hls.go is
+// independent of what we measure here.
+var softwareBenchmarkRungs = []benchmarkRung{
+	{height: 1080, width: 1920},
+	{height: 720, width: 1280},
+	{height: 480, width: 854},
+}
+
+// realtimeMarginSoftware is how much faster than realtime a synthetic encode
+// must run before we call a rung "sustainable". 1.5× leaves headroom for two
+// things the benchmark does NOT measure: (a) decoding the real source —
+// software HEVC / 10-bit decode is heavier than encoding the synthetic clip —
+// and (b) real content being busier than testsrc2 (which x264 compresses
+// faster than film grain or motion).
+const realtimeMarginSoftware = 1.5
+
+// benchmarkClipSeconds is the synthetic clip length. Short enough that a
+// capable host finishes the 1080p rung in well under a second, long enough to
+// average out process spin-up.
+const benchmarkClipSeconds = 3
+
+// BenchmarkMaxTranscodeHeight returns the largest output height this host can
+// software-transcode in real time, one of {1080,720,480}. Hardware encoders
+// return 2160 WITHOUT benchmarking — NVENC/QSV/VAAPI/VideoToolbox all sustain
+// 4K and a probe would only add startup latency.
+//
+// The point is the weak end. A low-power NAS or an old CPU can be
+// ffmpeg-capable yet unable to keep up with a 1080p software encode, so the
+// historical static 1080 ceiling makes the web side attempt a transcode that
+// stutters. Measuring real throughput lets decideStreamPlan route oversized
+// sources to an external player instead. Floors at 480: a box that can't
+// sustain even that is barely functional, and 480-or-smaller sources transcode
+// cheaply regardless — anything larger is already gated out by the 480 ceiling.
+func BenchmarkMaxTranscodeHeight(ctx context.Context, ffmpegPath string, hw HWAccel) int {
+	if hw != HWAccelNone {
+		return 2160
+	}
+	if ffmpegPath == "" {
+		return 1080 // no benchmark possible; keep the historical default
+	}
+	for _, rung := range softwareBenchmarkRungs {
+		factor, ok := measureEncodeRealtimeFactor(ctx, ffmpegPath, rung)
+		if !ok {
+			// Probe couldn't run (timeout / exec error) — try a lighter rung
+			// rather than treat the failure as a measured "fast enough".
+			log.Printf("[transcode] encode benchmark: %dp probe failed — trying lower", rung.height)
+			continue
+		}
+		if factor >= realtimeMarginSoftware {
+			log.Printf("[transcode] encode benchmark: software ceiling %dp (%.1f× realtime)", rung.height, factor)
+			return rung.height
+		}
+		log.Printf("[transcode] encode benchmark: %dp only %.1f× realtime (<%.1f×) — trying lower", rung.height, factor, realtimeMarginSoftware)
+	}
+	log.Printf("[transcode] encode benchmark: host can't sustain 480p software encode — flooring ceiling at 480 (oversized sources route to external)")
+	return 480
+}
+
+// measureEncodeRealtimeFactor encodes benchmarkClipSeconds of synthetic video
+// at the rung's resolution using the real streaming encoder settings (libx264
+// superfast, no B-frames) to /dev/null and returns clipDuration/wallTime — the
+// realtime factor. ok=false when the probe couldn't run, so the caller skips
+// rather than treating the failure as a fast result. Each probe is bounded so
+// a wedged ffmpeg can't stall daemon startup.
+func measureEncodeRealtimeFactor(ctx context.Context, ffmpegPath string, rung benchmarkRung) (float64, bool) {
+	// A 3 s superfast encode that takes longer than 12 s is <0.25× realtime —
+	// already far below the 1.5× bar — so capping here only kills genuinely
+	// hopeless rungs early and keeps worst-case startup bounded.
+	bctx, cancel := context.WithTimeout(ctx, 12*time.Second)
+	defer cancel()
+
+	size := strconv.Itoa(rung.width) + "x" + strconv.Itoa(rung.height)
+	args := []string{
+		"-hide_banner", "-nostats", "-loglevel", "error",
+		"-f", "lavfi",
+		"-i", "testsrc2=size=" + size + ":rate=24:duration=" + strconv.Itoa(benchmarkClipSeconds),
+		"-c:v", "libx264", "-preset", "superfast", "-threads", "0",
+		"-bf", "0", "-sc_threshold", "0",
+		"-f", "null", "-",
+	}
+	start := time.Now()
+	err := exec.CommandContext(bctx, ffmpegPath, args...).Run()
+	elapsed := time.Since(start)
+	if err != nil || elapsed <= 0 {
+		return 0, false
+	}
+	return float64(benchmarkClipSeconds) / elapsed.Seconds(), true
+}

internal/engine/encode_benchmark_test.go

@@ -0,0 +1,52 @@
+package engine
+
+import (
+	"context"
+	"os/exec"
+	"testing"
+)
+
+func TestBenchmarkMaxTranscodeHeight_HardwareSkipsProbe(t *testing.T) {
+	// Hardware encoders return 2160 without touching ffmpeg — pass a bogus path
+	// to prove no subprocess runs.
+	for _, hw := range []HWAccel{HWAccelNVENC, HWAccelQSV, HWAccelVAAPI, HWAccelVideoToolbox} {
+		got := BenchmarkMaxTranscodeHeight(context.Background(), "/nonexistent/ffmpeg", hw)
+		if got != 2160 {
+			t.Errorf("hw=%s: got %d, want 2160", hw, got)
+		}
+	}
+}
+
+func TestBenchmarkMaxTranscodeHeight_NoFFmpegKeepsDefault(t *testing.T) {
+	if got := BenchmarkMaxTranscodeHeight(context.Background(), "", HWAccelNone); got != 1080 {
+		t.Errorf("empty ffmpeg path: got %d, want 1080 (historical default)", got)
+	}
+}
+
+func TestBenchmarkMaxTranscodeHeight_SoftwareReturnsValidRung(t *testing.T) {
+	ffmpeg, err := exec.LookPath("ffmpeg")
+	if err != nil {
+		t.Skip("ffmpeg not on PATH — software benchmark needs a real encoder")
+	}
+	got := BenchmarkMaxTranscodeHeight(context.Background(), ffmpeg, HWAccelNone)
+	switch got {
+	case 1080, 720, 480:
+		// any rung is valid; the exact one depends on the host's CPU.
+	default:
+		t.Errorf("software ceiling = %d, want one of {1080,720,480}", got)
+	}
+}
+
+func TestMeasureEncodeRealtimeFactor_RealEncoder(t *testing.T) {
+	ffmpeg, err := exec.LookPath("ffmpeg")
+	if err != nil {
+		t.Skip("ffmpeg not on PATH")
+	}
+	factor, ok := measureEncodeRealtimeFactor(context.Background(), ffmpeg, benchmarkRung{height: 480, width: 854})
+	if !ok {
+		t.Fatal("480p probe failed to run on a host with ffmpeg")
+	}
+	if factor <= 0 {
+		t.Errorf("realtime factor = %.2f, want > 0", factor)
+	}
+}