lvqr-transcode 1.0.0

Server-side transcoding for LVQR (Tier 4 item 4.6).

Generates an ABR ladder (720p / 480p / 240p by default) from a single high-resolution source broadcast, with the output renditions re-injected into the caller-supplied [lvqr_fragment::FragmentBroadcasterRegistry] under <source>/<rendition> broadcast names. Every egress surface (LL-HLS, DASH, MoQ relay, archive) picks them up without per-protocol wiring; the LL-HLS master playlist composer emits one #EXT-X-STREAM-INF per rendition automatically.

What this crate ships

Always available (default features):

• [Transcoder] trait + [TranscoderFactory] + [TranscoderContext] -- the "subscribe / drain / panic-isolate" shape generalised from lvqr_agent::Agent. Each factory carries its own [RenditionSpec].
• [TranscodeRunner] + [TranscodeRunnerHandle] -- registry-side installer + cheaply-cloneable handle, mirroring lvqr_agent::AgentRunner.
• [RenditionSpec] with width / height / bitrate fields, [RenditionSpec::preset_720p] / preset_480p / preset_240p constructors, and [RenditionSpec::default_ladder].
• [PassthroughTranscoder] -- in-memory observer that proves the registry-callback / drain / panic-isolation wiring; useful for tests and as a metrics tap.
• [AudioPassthroughTranscoder] -- copies <source>/1.mp4 audio fragments verbatim into every rendition's audio track so each rendition broadcaster is a self-contained mp4 the LL-HLS bridge drains without special-casing the missing audio.

Behind the transcode feature (pulls gstreamer-rs 0.23 + base/good/bad/ugly + gst-libav from the host):

• [SoftwareTranscoder] / [SoftwareTranscoderFactory] -- the appsrc -> qtdemux -> h264parse -> avdec_h264 -> videoscale -> x264enc -> ... -> mp4mux -> appsink ladder, one worker thread per (source, rendition) pair, bounded mpsc.
• [AacToOpusEncoder] / [AacToOpusEncoderFactory] -- AAC -> Opus transcoder used by lvqr-whep (under its aac-opus feature) so AAC publishers reach Opus-negotiated WHEP subscribers.

Behind one of the hw-* features (each implies transcode):

• hw-videotoolbox -- [VideoToolboxTranscoder] / [VideoToolboxTranscoderFactory] for macOS via Apple's vtenc_h264_hw (the applemedia plugin from gst-plugins-bad).
• hw-nvenc -- [NvencTranscoder] / [NvencTranscoderFactory] for Linux + Nvidia GPUs via nvh264enc (the nvcodec plugin from gst-plugins-bad, driven by the CUDA runtime).
• hw-vaapi -- [VaapiTranscoder] / [VaapiTranscoderFactory] for Linux + Intel iGPU / AMD via vah264enc (the modern va plugin from gst-plugins-bad, superseding the deprecated vaapih264enc from gstreamer-vaapi).
• hw-qsv -- [QsvTranscoder] / [QsvTranscoderFactory] for Linux + Intel Quick Sync via qsvh264enc (the qsv plugin from gst-plugins-bad, driving Intel Media SDK / oneVPL).

All four HW backends mirror SoftwareTranscoderFactory shape verbatim -- same Transcoder trait, same lifecycle, same <source>/<rendition> output broadcast naming, same bounded mpsc + dedicated worker thread per (source, rendition) pair -- and only swap the GStreamer encoder element + property mapping. HW-only path is intentional across all four: a factory that silently falls back to CPU encoding under load defeats the point of an operator-pickable hardware tier. Each factory's is_available() probes the required encoder element at construction and build() opts out cleanly with a warn log when missing.

Future sessions may extract the shared scaffolding into a dedicated pipeline.rs module (per the "three is the threshold for an abstraction" rule). The current shape is intentional code duplication: each backend stays readable on its own and the cost of cross-backend changes is small enough that the mechanical- sharing tradeoff is not yet a win.

Where this crate fits in the consumer family

Pattern-matches the existing [lvqr_fragment::FragmentBroadcasterRegistry] consumers:

Operator wiring

lvqr-cli exposes --transcode-rendition 720p,480p,240p (or a .toml RenditionSpec path) and, on hw-videotoolbox builds, --transcode-encoder software|videotoolbox. End-to-end shape: ingest one source RTMP stream, the LL-HLS master playlist advertises one variant per rendition + the source.