Per Frontend Spec (A2.4 pre-merge note): MAX_UPLOAD_SIZE was hardcoded to
100 MB and would force-compress 100-300 MB files that should skip. Reads
from VIDEO_MAX_SIZE now. Pulled the target/passthrough bitrate and max
dimension out of compress.ts into src/lib/media/video/constants.ts so the
compression contract lives in one place without bloating the global
constants file.
- Android non-AAC audio is now transcoded to AAC (matching iOS) via a
source-decoder -> AAC-encoder pre-pass that captures the encoder's
output format before the muxer starts, then writes buffered samples
after the video pipeline finishes. Falls back to dropping audio if
the transcode fails.
- compress.ts wraps probe() in try/catch and falls through to
passthrough on failure instead of throwing.
- Probers expose isHDR (HLG/PQ via color transfer, plus Dolby Vision
codecs/mimes); shouldCompress forces compression for HDR sources so
the SDR BT.709 path always runs. Also sets
KEY_COLOR_TRANSFER_REQUEST=SDR on the Android decoder (API 31+) so
HDR sources tone-map to SDR pixels instead of being mislabeled.
N1: Delete unused CodecSelector.selectEncoder. VideoCompressor calls
findEncoder directly; selectEncoder was dead since the initial commit.
N2: Switch Android output filename from System.currentTimeMillis() to
UUID.randomUUID(). Eliminates the collision risk when two compressions
start in the same millisecond.
N3: Guard CheckedContinuation against double-resume in both
processVideoTrack and processAudioTrack. AVAssetWriter shouldn't
re-invoke the requestMediaDataWhenReady block after markAsFinished, but
if it did, the previous code would crash via CheckedContinuation's
double-resume detection. Added a `finished` flag plus a local `finish`
closure that no-ops on second call.
N4: DataRateLimits window duration changed from Int `1` to Double `1.0`.
VideoToolbox expects the duration value as a CFNumber with float
semantics; Int bridges to NSNumber(int) which works on most iOS
versions but is not spec-correct.
B1: Android encoder format now sets KEY_COLOR_STANDARD = BT709,
KEY_COLOR_TRANSFER = SDR_VIDEO, KEY_COLOR_RANGE = LIMITED (API 24+).
Previously the encoder inherited or emitted default color metadata,
which meant HDR sources produced incorrectly-tagged output on Android.
iOS already had AVVideoColorPropertiesKey set correctly.
B2: Replace single currentCompressor reference with a per-job map
keyed by jobId. The cancel function now takes a jobId and only cancels
that specific job. Prevents the previous race where a second compress
call overwrote the reference and made the first job uncancellable.
B3: Clamp frameRateCap to >= 1 at the native module boundary. Previously
a value of 0 from JS would cause divide-by-zero (Android Long division
ArithmeticException, iOS CMTime Infinity / Int32(0) frameDuration trap).
Replaces react-native-compressor's video path with a local Expo module.
Native pipeline:
- iOS: AVAssetReader + AVMutableVideoComposition + AVAssetWriter with
VideoToolbox encode, BGRA reader for HDR -> SDR conversion, BT.709
color tagging, DataRateLimits hard cap, AAC re-encode.
- Android: MediaExtractor -> MediaCodec(decoder) -> GL pipeline ->
MediaCodec(encoder) -> MediaMuxer. BITRATE_MODE_CBR for tight target
enforcement, GL transform-matrix rotation (no double-rotation),
raw AAC passthrough, hardware encoder selection with software
fallback, QTI AVC denylist.
Codec selection: 'auto' resolves to h264 (HLS pipeline + licensing).
HEVC remains opt-in via codec: 'hevc' for future feature-flagged use.
compress.ts adds probe-based smart-skip: clips that are already at or
below 5 Mbps / 1920px / 100MB bypass re-encoding entirely.
react-native-compressor stays as a dep for now since pickVideo and
VideoTranscodeBackdrop still use its non-compression helpers; full
removal is a follow-up.
vineyardbovines