Do not forward the audio codec config buffer as an encoded frame

[RomanHerbstmann]

Problem

MediaCodec emits the audio codec configuration as its first output buffer, flagged with BUFFER_FLAG_CODEC_CONFIG. For AAC that buffer is the 2-byte AudioSpecificConfig.

BaseEncoder.processOutput forwards every buffer that checkBuffer accepts:

if (checkBuffer(byteBuffer, bufferInfo)) sendBuffer(byteBuffer, bufferInfo);

VideoEncoder.checkBuffer keeps SPS/PPS out of that path, but AudioEncoder.checkBuffer only validates the timestamp, so the config buffer is forwarded like a regular encoded frame. It then reaches

• the senders, which transmit it as a tiny bogus AAC frame, and
• recordAudio → AndroidMuxerRecordController.write → MediaMuxer.writeSampleData, which stores it as a media sample.

Impact

Recorded MP4 files start with a 2-byte audio sample that is identical to the track's extradata. FFmpeg and VLC skip it with Input buffer exhausted before END element found, but Chrome aborts playback of the entire file:

PIPELINE_ERROR_DECODE: Failed to send audio packet for decoding:
{timestamp=0 duration=103084 size=2 is_key_frame=1 encrypted=0}

Verified on a 1.4 GB, 66-minute recording produced from an RTMP stream: the first audio packet is 12 10 (AAC-LC, 44.1 kHz, stereo), byte for byte the same as the track's extradata. Remuxing the same file with that single packet dropped makes it play in Chrome; nothing else changes.

Fix

Reject buffers carrying BUFFER_FLAG_CODEC_CONFIG in AudioEncoder.checkBuffer, mirroring how video keeps SPS/PPS out of the frame path.

This is safe because the configuration is derived independently on every path:

• RTMP: AacPacket.sendAudioInfo(sampleRate, isStereo, codec) builds AacAudioSpecificConfig from sample rate and channel count.
• RTSP: SdpBody signals it from sample rate and channel count.
• Recording: the muxer track is created from the MediaFormat passed to setAudioFormat in onFormatChanged, which already carries csd-0.

The change covers every MediaCodec-based audio codec (AAC, HE-AAC, Opus). G711 uses the software path in BaseEncoder.processG711() and has no codec-config buffer, so it is unaffected. The filter is deliberately kept in AudioEncoder rather than BaseEncoder, because video still relies on the config buffer for manual SPS/PPS extraction when formatChanged arrives late.

Tests

Added encoder/src/test/java/com/pedro/encoder/AudioEncoderCheckBufferTest.kt with two cases: a buffer flagged as codec config is rejected, a normal frame with a valid timestamp is accepted. ./gradlew :encoder:testDebugUnitTest passes 24/24.

Note for anyone writing further tests in this module: it sets unitTests.isReturnDefaultValues = true, so MediaCodec.BufferInfo.set(...) is a no-op stub and the fields have to be assigned directly.