Tab recording in 0.9.0: h264_nvenc segfaults on input aspect change, and seed screenshot can time out

Since 0.9.0, tab recording (recording.start) fails on my machine in two independent ways. 0.8.0 (which used chrome.tabCapture) worked.

Environment

• playwriter CLI 0.9.0, extension 0.0.170
• Linux x64 (Ubuntu 22.04), Node 24
• ffmpeg 4.4.2 (distro package)
• NVIDIA T1000, driver 550.120

1. ffmpeg segfaults with h264_nvenc when the input frame aspect ratio changes

For MP4 output the recorder picks the encoder from detectEncoder(), which chooses h264_nvenc here. Recording fails with:

Error: ffmpeg exited with SIGSEGV

I can reproduce it without Chrome. Two MJPEG frames whose aspect ratio differs (1920x992, then 1440x810), piped into the same arguments the recorder builds:

ffmpeg -hide_banner -loglevel error -nostdin -f image2pipe -c:v mjpeg \
  -framerate 30 -thread_queue_size 512 -i two-frames.mjpeg \
  -vf 'scale=1920:992:force_original_aspect_ratio=decrease,pad=1920:992:(ow-iw)/2:(oh-ih)/2,format=yuv420p' \
  -c:v h264_nvenc -b:v 5714k -movflags +faststart -y out.mp4

gdb puts the top frames inside libnvidia-encode.so, called from avcodec_send_frame. The kernel log shows the live recording crash and the reproduction crashing at the same offset. So it may be an NVIDIA or ffmpeg 4.4 bug, but the recorder is what sends it frames that change size. The single-frame probe in detectEncoder() cannot catch this.

2. The first screenshot can time out before ffmpeg starts

Separately, recording.start sometimes fails before ffmpeg is spawned:

cdpSession.send: Protocol error (Page.captureScreenshot):
Extension request timeout after 30000ms: forwardCDPCommand

The stack is PageCapture.start → ScreenRecording.start → recording.start, on the initial Page.captureScreenshot. Raising the CLI --timeout does not help, because the 30 s limit is inside the extension. If I call Page.startScreencast on the same CDP session first, the screenshot returns in about 3 s. That worked in 5 of 5 recordings.

Workaround

There is no codec option on recording.start, and no env override. I work around both problems by using the exported PageCapture with a buildArgs callback that calls buildRecordingFfmpegArgs({ ..., codec: 'libx264' }), and by starting the screencast before recording.

Suggestions

• Let recording.start accept a codec (or honour an env var), or fall back to libx264 when the hardware encoder exits abnormally.
• Or keep the output size constant before encoding, so the encoder never sees a change in input size or aspect.
• Start the screencast before taking the first screenshot, or don't block on that screenshot.