fix(helper): NVENC scale_cuda filter syntax and NVDEC reference-frame pool

Two NVENC pipeline issues were dropping every Tubi-class (HEVC source)
segment back to the software encoder fallback:

scale_cuda filter chain. The previous filter string
"scale_cuda=W:H:force_original_aspect_ratio=decrease,format=nv12" had
a comma before format=nv12, so ffmpeg parsed it as two filters:
scale_cuda (GPU surface output) followed by a standalone format=nv12
CPU-side filter. The auto_scale ffmpeg inserted to bridge GPU->CPU
couldn't convert the surface and surfaced as
"Impossible to convert between the formats supported by the filter
'Parsed_scale_cuda_0' and the filter 'auto_scale_0'". Move format
inside scale_cuda's own options
("scale_cuda=w=W:h=H:format=nv12:force_original_aspect_ratio=decrease")
so the whole filter stays on GPU and feeds h264_nvenc directly.

NVDEC reference-frame pool. HEVC sources with longer reference chains
than the default extra_hw_frames pool covers ("Could not find ref with
POC N") would drop frames mid-decode and produce a short or empty
segment that the upload validator then rejected. Bump -extra_hw_frames
to 16 on the NVENC backend so the pool keeps the reference frames the
decoder still needs.

Author: RealWhyKnot

src/WKVRCProxy.Tests/FfmpegHelperTests.cs

@@ -153,7 +153,7 @@ public void TranscodeWorkerProcess_BuildsSafeNvencSegmentCommand()
         Assert.Contains(command.Arguments, argument => argument.Contains("Mozilla/5.0 (Windows NT 10.0; Win64; x64)", StringComparison.Ordinal));
         Assert.Contains("h264_nvenc", command.Arguments);
         AssertHardwareDecodeBeforeInput(command.Arguments, "cuda", "cuda");
-        Assert.Contains(command.Arguments, argument => argument.StartsWith("scale_cuda=1280:720:", StringComparison.Ordinal));
+        Assert.Contains(command.Arguments, argument => argument.StartsWith("scale_cuda=w=1280:h=720:", StringComparison.Ordinal));
         Assert.Contains(command.Arguments, argument => argument.Contains("format=nv12", StringComparison.Ordinal));
         Assert.Contains("-force_key_frames", command.Arguments);
         Assert.Contains("-c:a", command.Arguments);

src/WKVRCProxy/Helper/TranscodeWorkerProcess.cs

@@ -245,7 +245,12 @@ private static IReadOnlyList<string> HardwareDecodeOptionsFor(HardwareEncoderCap
     {
         return encoder.Backend switch
         {
-            HardwareEncoderBackend.Nvenc => new[] { "-hwaccel", "cuda", "-hwaccel_output_format", "cuda" },
+            // -extra_hw_frames widens NVDEC's reference-frame pool. Some HEVC sources
+            // (Tubi's catalog among them) use longer reference chains than the cuvid
+            // default 25 covers and surface as "Could not find ref with POC N" decoder
+            // warnings followed by a black/short segment. Bumping the pool stops the
+            // decoder from dropping reference frames it still needs.
+            HardwareEncoderBackend.Nvenc => new[] { "-hwaccel", "cuda", "-hwaccel_output_format", "cuda", "-extra_hw_frames", "16" },
             HardwareEncoderBackend.Qsv => new[] { "-hwaccel", "qsv", "-hwaccel_output_format", "qsv" },
             HardwareEncoderBackend.Amf or HardwareEncoderBackend.MediaFoundation =>
                 new[] { "-hwaccel", "d3d11va", "-hwaccel_output_format", "d3d11" },
@@ -259,8 +264,15 @@ private static string HardwareScaleFilterFor(HardwareEncoderCapability encoder,
         string height = safeHeight.ToString(CultureInfo.InvariantCulture);
         return encoder.Backend switch
         {
+            // `format=nv12` must be part of scale_cuda's option list (colon-
+            // separated), not a separate filter. The previous form used a comma,
+            // which made ffmpeg parse `format=nv12` as a standalone CPU-side
+            // filter; an auto_scale would then be inserted to bridge the CUDA
+            // surface from scale_cuda to the CPU format filter and fail with
+            // "Impossible to convert between the formats supported by the
+            // filter 'Parsed_scale_cuda_0' and the filter 'auto_scale_0'".
             HardwareEncoderBackend.Nvenc =>
-                "scale_cuda=" + width + ":" + height + ":force_original_aspect_ratio=decrease,format=nv12",
+                "scale_cuda=w=" + width + ":h=" + height + ":format=nv12:force_original_aspect_ratio=decrease",
             HardwareEncoderBackend.Qsv =>
                 "scale_qsv=w=" + width + ":h=" + height + ":format=nv12",
             HardwareEncoderBackend.Amf or HardwareEncoderBackend.MediaFoundation =>