feat(stream): GPU HDR tonemap via libplacebo

Prefer the single-pass Vulkan libplacebo filter over the CPU zscale chain
for HDR->SDR tonemapping when the agent ffmpeg has it. One GPU pass does
tonemap + BT.709 primaries/transfer/matrix + 8-bit yuv420p, replacing the
four-stage zscale chain and its trailing format=/setparams. Higher quality,
far cheaper than the CPU path, and present on builds that lack zscale.

- FFmpegSupportsLibplacebo probe (cached, mirrors FFmpegSupportsZscale)
- HasLibplacebo on TranscodeRuntime, wired from buildTranscodeRuntime
- hls.go: videoTail picks libplacebo when present (not h264_vaapi), else
  keeps the zscale tonemap + format chain
- test: libplacebo replaces the zscale chain, never runs alongside it

internal/cmd/player_session_registry.go

@@ -95,5 +95,8 @@ func buildTranscodeRuntime(ctx context.Context, cfg config.Config) engine.Transc
 		// Tonemap HDR→SDR only when this ffmpeg build has zscale; otherwise the
 		// filter would error and break playback, so HDR plays untonemapped.
 		TonemapHDR: engine.FFmpegSupportsZscale(ffmpegPath),
+		// libplacebo (GPU) is preferred over zscale when present — checked here so
+		// the per-session arg builder can pick it for HDR sources.
+		HasLibplacebo: engine.FFmpegSupportsLibplacebo(ffmpegPath),
 	}
 }

internal/engine/hls.go

@@ -1353,26 +1353,37 @@ func buildHLSFFmpegArgsAt(cfg HLSSessionConfig, probe *StreamProbe, tmpDir strin
 		hwUploadTail = ",hwupload"
 		colorTail = ""
 	}
-	// HDR→SDR tonemap, inserted after the scale (downscale-first = fewer pixels
+	// HDR→SDR tonemap, after the scale (downscale-first = fewer pixels to map).
-	// to tonemap) and before format=. Only for an HDR source on a zscale-capable
+	// Prefer libplacebo (GPU, ONE pass — it also emits the BT.709 colorspace +
-	// ffmpeg; the trailing comma in hdrTonemapChain slots it in front of format=.
+	// 8-bit format, so it REPLACES the format/setparams tail); else the zscale
+	// CPU chain; else play untonemapped (desaturated, last resort). Skip
+	// libplacebo on VAAPI: its Vulkan surface flow doesn't compose with our
+	// nv12+hwupload path, so VAAPI keeps the zscale-or-none behaviour.
+	useLibplacebo := probe.HDR != "" && cfg.Transcode.HasLibplacebo && codec != "h264_vaapi"
 	tonemap := ""
-	if probe.HDR != "" && cfg.Transcode.TonemapHDR {
+	if probe.HDR != "" && cfg.Transcode.TonemapHDR && !useLibplacebo {
 		tonemap = hdrTonemapChain
 	}
-	// Core video chain (scale + optional tonemap + pixel format + color metadata),
+	// videoTail = everything after the scale: either libplacebo (tonemap +
-	// WITHOUT the optional hwUploadTail — that has to run last, after any subtitle
+	// colorspace + format in one) or the (optional zscale) tonemap then the
-	// overlay, so it's appended separately below.
+	// format + color-metadata tail. No leading comma — the scale chain ends in one.
+	videoTail := tonemap + "format=" + pixFormat + colorTail
+	if useLibplacebo {
+		videoTail = libplaceboTonemapFilter
+	}
+	// Core video chain (scale + tonemap/format tail), WITHOUT the optional
+	// hwUploadTail — that has to run last, after any subtitle overlay, so it's
+	// appended separately below.
 	var vchain string
 	if maxH > 0 && probe.Height > maxH {
 		vchain = fmt.Sprintf(
-			"scale=-2:%d:force_original_aspect_ratio=decrease,scale=trunc(iw/2)*2:trunc(ih/2)*2,%sformat=%s%s",
+			"scale=-2:%d:force_original_aspect_ratio=decrease,scale=trunc(iw/2)*2:trunc(ih/2)*2,%s",
-			maxH, tonemap, pixFormat, colorTail,
+			maxH, videoTail,
 		)
 	} else {
 		vchain = fmt.Sprintf(
-			"scale=trunc(iw/2)*2:trunc(ih/2)*2,%sformat=%s%s",
+			"scale=trunc(iw/2)*2:trunc(ih/2)*2,%s",
-			tonemap, pixFormat, colorTail,
+			videoTail,
 		)
 	}
 	if burnIdx >= 0 {

internal/engine/tonemap.go

@@ -28,11 +28,53 @@ import (
 // possible follow-up if HLG/DV-only sources become common.
 const hdrTonemapChain = "zscale=t=linear:npl=100,format=gbrpf32le,zscale=p=bt709,tonemap=tonemap=hable:desat=0,zscale=t=bt709:m=bt709:r=tv,"
 
+// libplaceboTonemapFilter maps an HDR source to SDR BT.709 in a SINGLE GPU pass
+// (Vulkan): tone-map the HDR curve, convert primaries/transfer/matrix to BT.709
+// limited range, and output 8-bit yuv420p — so it REPLACES the zscale chain AND
+// the trailing `format=yuv420p,setparams=bt709` (it does both). Higher quality
+// and far cheaper than the CPU zscale chain, and the agent's ffmpeg has it where
+// zscale is missing. It does NOT scale here — the CPU scale chain runs first
+// (it owns the even-dimension rounding libx264/nvenc require). No trailing comma:
+// it's the last filter in the chain.
+const libplaceboTonemapFilter = "libplacebo=colorspace=bt709:color_primaries=bt709:color_trc=bt709:range=tv:format=yuv420p:tonemapping=bt.2390"
+
 var (
 	zscaleCacheMu sync.Mutex
 	zscaleCache   = map[string]bool{}
+
+	libplaceboCacheMu sync.Mutex
+	libplaceboCache   = map[string]bool{}
 )
 
+// FFmpegSupportsLibplacebo reports whether the ffmpeg binary has the libplacebo
+// filter (Vulkan GPU HDR tonemap + colorspace). Preferred over zscale when both
+// exist. Cached per path; a probe failure is treated as "no". Mirrors
+// FFmpegSupportsZscale.
+func FFmpegSupportsLibplacebo(ffmpegPath string) bool {
+	if ffmpegPath == "" {
+		return false
+	}
+	libplaceboCacheMu.Lock()
+	if v, ok := libplaceboCache[ffmpegPath]; ok {
+		libplaceboCacheMu.Unlock()
+		return v
+	}
+	libplaceboCacheMu.Unlock()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	out, err := exec.CommandContext(ctx, ffmpegPath, "-hide_banner", "-filters").Output()
+	supported := err == nil && bytes.Contains(out, []byte("libplacebo"))
+
+	libplaceboCacheMu.Lock()
+	libplaceboCache[ffmpegPath] = supported
+	libplaceboCacheMu.Unlock()
+	if supported {
+		log.Printf("[tonemap] ffmpeg has libplacebo — HDR sources tonemapped on the GPU (preferred)")
+	}
+	return supported
+}
+
 // FFmpegSupportsZscale reports whether the ffmpeg binary at path was built with
 // the zscale filter (libzimg), required for HDR→SDR tonemapping. Cached per
 // path. A detection failure (binary missing, exec error) is treated as "no" so

internal/engine/tonemap_test.go

@@ -66,6 +66,25 @@ func TestTonemap_AppliedInNoDownscaleBranch(t *testing.T) {
 	}
 }
 
+func TestTonemap_LibplaceboPreferredOverZscale(t *testing.T) {
+	// HDR source + an ffmpeg with libplacebo → the single GPU filter replaces
+	// the whole CPU zscale chain (and the trailing format=/setparams it folds in).
+	cfg := HLSSessionConfig{
+		SessionID:  "test",
+		SourcePath: "/movies/x.mkv",
+		Quality:    "720p",
+		Transcode:  TranscodeRuntime{FFmpegPath: "/usr/bin/ffmpeg", HWAccel: HWAccelNone, TonemapHDR: true, HasLibplacebo: true},
+	}
+	probe := &StreamProbe{Width: 3840, Height: 2160, BitDepth: 10, HDR: "HDR10", DurationSec: 100}
+	vf := vfChain(strings.Join(buildHLSFFmpegArgsAt(cfg, probe, "/tmp/t", 0, 0), " "))
+	if !strings.Contains(vf, "libplacebo") {
+		t.Fatalf("libplacebo-capable ffmpeg: expected libplacebo filter, got %q", vf)
+	}
+	if strings.Contains(vf, "zscale=t=linear") || strings.Contains(vf, "tonemap=tonemap=hable") {
+		t.Errorf("libplacebo must replace the zscale chain, not run alongside it: %q", vf)
+	}
+}
+
 func TestTonemap_SkippedWhenFFmpegLacksZscale(t *testing.T) {
 	vf := vfChain(hlsArgsFor("HDR10", false, HWAccelNone))
 	if strings.Contains(vf, "zscale") || strings.Contains(vf, "tonemap") {

internal/engine/transcode_quality.go

@@ -18,6 +18,10 @@ type TranscodeRuntime struct {
 	// Set only when the ffmpeg build has zscale (FFmpegSupportsZscale); without
 	// it the tonemap filter would error and break playback, so it stays off.
 	TonemapHDR bool
+	// HasLibplacebo: the ffmpeg build has the libplacebo filter (GPU HDR tonemap).
+	// Preferred over the zscale chain for HDR sources — one GPU pass, higher
+	// quality, and present where zscale is missing.
+	HasLibplacebo bool
 }
 
 // qualityCap maps a session's Quality label to a (MaxHeight, VideoBitrate)