fix(engine): accept libx264 preset names with NVENC and QSV (#442)

NVENC rejects the libx264 preset vocabulary (ultrafast / medium / slow /
...) with AVERROR(EINVAL) ("Error applying encoder options: Invalid
argument"), which surfaces as a bare `FFmpeg exited with code -22` from
spawn(). Because ENCODER_PRESETS passes these names straight through to
h264_nvenc / hevc_nvenc, every `--gpu` render using the `draft` tier
failed; `standard` (medium) and `high` (slow) only worked coincidentally
on ffmpeg builds that happened to accept those aliases. QSV has the same
problem on a narrower set (ultrafast / superfast / placebo).

Add `mapPresetForGpuEncoder` in utils/gpuEncoder.ts that translates the
libx264 vocabulary to each encoder's native names:

- nvenc: libx264 -> p1..p7 (already-native pN values pass through);
  unknown values fall back to p4 (medium)
- qsv:   ultrafast / superfast -> veryfast; placebo -> veryslow;
  everything else passes through
- videotoolbox / vaapi / null: unchanged

Both buildEncoderArgs (chunkEncoder.ts) and buildStreamingArgs
(streamingEncoder.ts) now route through the helper before pushing
`-preset` to the ffmpeg arg vector.

To make the next encoder-options failure diagnosable without re-running
ffmpeg by hand, \`formatFfmpegError\` in utils/runFfmpeg.ts now appends
the last 15 non-empty stderr lines to the error string. The four call
sites that previously swallowed stderr (encodeFramesFromDir,
muxVideoWithAudio, applyFaststart, and the streaming encoder exit
handler) have been updated.

Tested end-to-end on an RTX 4080 with ffmpeg 8.1 NVENC across
\`--quality draft|standard|high\` plus \`--video-bitrate\` and \`--crf\`
overrides; the 6 renders were visually equivalent to the CPU baseline.

Co-authored-by: roi32 <75878108+roi32@users.noreply.github.com>

packages/engine/src/utils/gpuEncoder.ts

@@ -56,3 +56,54 @@ export function getGpuEncoderName(encoder: GpuEncoder, codec: "h264" | "h265"):
       return codec === "h264" ? "libx264" : "libx265";
   }
 }
+
+// libx264 preset names (ultrafast/superfast/.../placebo) mapped to the
+// equivalent NVENC p1..p7 preset. NVENC rejects libx264 names with
+// AVERROR(EINVAL) ("Error applying encoder options: Invalid argument"),
+// which surfaces as a generic "FFmpeg exited with code -22" — so callers
+// that share a single `preset` field across CPU and GPU paths (e.g. the
+// `draft`/`standard`/`high` quality tiers) must translate before passing
+// the value to h264_nvenc / hevc_nvenc.
+const NVENC_PRESET_MAP: Record<string, string> = {
+  ultrafast: "p1",
+  superfast: "p1",
+  veryfast: "p2",
+  faster: "p3",
+  fast: "p4",
+  medium: "p4",
+  slow: "p5",
+  slower: "p6",
+  veryslow: "p7",
+  placebo: "p7",
+};
+
+// QSV accepts most libx264 preset names but rejects `ultrafast`,
+// `superfast`, and `placebo`. Map those to the nearest supported values.
+const QSV_PRESET_MAP: Record<string, string> = {
+  ultrafast: "veryfast",
+  superfast: "veryfast",
+  placebo: "veryslow",
+};
+
+/**
+ * Translate a libx264-style `-preset` value to one accepted by the given
+ * GPU encoder.
+ *
+ * - `nvenc`: libx264 names → `p1`..`p7`. Already-native `pN` values pass
+ *   through unchanged. Unknown values fall back to `p4` (medium).
+ * - `qsv`:  `ultrafast`/`superfast`/`placebo` → nearest supported name;
+ *   everything else passes through.
+ * - `videotoolbox`, `vaapi`, `null`: no remap (they either ignore `-preset`
+ *   entirely or accept the libx264 vocabulary).
+ */
+export function mapPresetForGpuEncoder(encoder: GpuEncoder, preset: string): string {
+  switch (encoder) {
+    case "nvenc":
+      if (/^p[1-7]$/.test(preset)) return preset;
+      return NVENC_PRESET_MAP[preset] ?? "p4";
+    case "qsv":
+      return QSV_PRESET_MAP[preset] ?? preset;
+    default:
+      return preset;
+  }
+}