arcly-stream 0.1.7

//! Shared RTP/RTCP parsing and NAL-codec (de)packetization — RFC 3550 framing
//! plus H.264 (RFC 6184) and H.265 (RFC 7798) payload formats.
//!
//! Gated behind the internal `_rtp` marker, pulled in by both [`rtsp`] and
//! [`webrtc`]. The two transports differ only in how RTP packets reach the
//! process (TCP-interleaved / UDP for RTSP, DTLS-SRTP for WebRTC); once a packet
//! is in hand, reassembling NAL units into an Annex-B access unit is identical,
//! so it lives here once.
//!
//! [`rtsp`]: crate::protocol::rtsp
//! [`webrtc`]: crate::protocol::webrtc
//!
//! # What it does
//!
//! - [`RtpHeader::parse`] decodes the fixed RTP header (RFC 3550 §5.1), honoring
//!   the CSRC count and the extension-header flag to locate the payload.
//! - [`H264Depacketizer`] / [`H265Depacketizer`] turn a sequence of RTP payloads
//!   into complete access units in Annex-B form. Each handles the three NALU
//!   packetization modes for its codec — single NAL, aggregation (STAP-A type 24
//!   / AP type 48), and fragmentation (FU-A type 28 / FU type 49) — emitting an
//!   access unit when the marker bit is set or the timestamp advances.
//! - [`RtpPacketizer`] performs the reverse for egress (e.g. WebRTC WHEP),
//!   selecting the H.264 or H.265 payload format.
//!
//! # What it does not do
//!
//! Jitter-buffer reordering and loss concealment are the caller's concern — the
//! depacketizer assumes in-order delivery (true for TCP-interleaved RTSP; for
//! UDP/SRTP a small reorder buffer should sit in front of it). It reports a
//! [`DepacketizeError::OutOfOrder`] gap so a handler can request a keyframe
//! (PLI/FIR) rather than emit a corrupt access unit.

use bytes::Bytes;

/// Annex-B start code prefixed to every reassembled NAL unit.
const ANNEXB_START: [u8; 4] = [0, 0, 0, 1];

/// A parsed RTP fixed header (RFC 3550 §5.1).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct RtpHeader {
    /// Payload type (7 bits) — identifies the codec/format binding from SDP.
    pub payload_type: u8,
    /// Marker bit. For H.264 it flags the last packet of an access unit.
    pub marker: bool,
    /// 16-bit sequence number, increments by one per packet (wraps).
    pub sequence: u16,
    /// 32-bit media timestamp in the payload's clock (90 kHz for H.264 video).
    pub timestamp: u32,
    /// Synchronization source identifier.
    pub ssrc: u32,
    /// Byte offset at which the payload begins (past CSRCs and any extension).
    pub payload_offset: usize,
}

impl RtpHeader {
    /// Parse the fixed header from the front of `buf`, returning the header and
    /// the payload offset. Returns `None` if `buf` is too short or the version
    /// field is not 2.
    pub fn parse(buf: &[u8]) -> Option<RtpHeader> {
        use super::byteops::ByteReader;
        let mut r = ByteReader::new(buf);
        let b0 = r.u8()?;
        if b0 >> 6 != 2 {
            return None; // RTP version must be 2
        }
        let has_extension = b0 & 0x10 != 0;
        let csrc_count = (b0 & 0x0F) as usize;
        let b1 = r.u8()?;
        let marker = b1 & 0x80 != 0;
        let payload_type = b1 & 0x7F;
        let sequence = r.u16_be()?;
        let timestamp = r.u32_be()?;
        let ssrc = r.u32_be()?;
        r.skip(csrc_count * 4)?; // CSRC list

        if has_extension {
            // Extension header: 2-byte profile id, 2-byte length (in 32-bit words).
            r.skip(2)?;
            let ext_words = r.u16_be()? as usize;
            r.skip(ext_words * 4)?;
        }
        Some(RtpHeader {
            payload_type,
            marker,
            sequence,
            timestamp,
            ssrc,
            payload_offset: r.position(),
        })
    }
}

/// Errors surfaced while depacketizing an RTP stream.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[non_exhaustive]
pub enum DepacketizeError {
    /// The packet was shorter than the format requires.
    Truncated,
    /// A sequence-number discontinuity was detected mid-fragment; the partial
    /// access unit was dropped. The handler should request a keyframe.
    OutOfOrder,
    /// An unsupported NAL/aggregation type was encountered.
    Unsupported(u8),
}

/// Depacketizes RFC 3640 AAC-hbr RTP payloads into raw AAC access units.
///
/// The common RTSP/SDP profile for AAC (`mode=AAC-hbr`, `sizelength=13`,
/// `indexlength=3`) frames each payload as a 2-byte **AU-headers-length** (in
/// bits), followed by one 2-byte AU-header per access unit (13-bit size +
/// 3-bit index), followed by the access units concatenated. One RTP packet may
/// carry several AAC frames; [`push`](Self::push) returns each as a separate
/// raw (ADTS-less) [`bytes::Bytes`].
#[derive(Debug, Clone, Copy, Default)]
pub struct AacDepacketizer {
    /// Bits per AU-header `size` field (13 for AAC-hbr).
    size_length: u8,
    /// Bits per AU-header `index`/`index-delta` field (3 for AAC-hbr).
    index_length: u8,
}

impl AacDepacketizer {
    /// A depacketizer for the standard AAC-hbr profile (`sizelength=13`,
    /// `indexlength=3`).
    pub fn new() -> Self {
        Self {
            size_length: 13,
            index_length: 3,
        }
    }

    /// A depacketizer with explicit AU-header field widths from the SDP `fmtp`.
    pub fn with_lengths(size_length: u8, index_length: u8) -> Self {
        Self {
            size_length,
            index_length,
        }
    }

    /// Split one RTP AAC-hbr payload into its constituent access units.
    pub fn push(&self, payload: &[u8]) -> Result<Vec<Bytes>, DepacketizeError> {
        if payload.len() < 2 {
            return Err(DepacketizeError::Truncated);
        }
        // Sizes wider than a 16-bit AU-header field are unsupported (and would
        // otherwise over-shift below). `with_lengths` can supply arbitrary widths.
        if self.size_length == 0 || self.size_length > 16 {
            return Err(DepacketizeError::Unsupported(self.size_length));
        }
        let header_bits = u16::from_be_bytes([payload[0], payload[1]]) as usize;
        let au_header_bits = self.size_length as usize + self.index_length as usize;
        if au_header_bits == 0 {
            return Err(DepacketizeError::Unsupported(0));
        }
        let header_bytes = header_bits.div_ceil(8);
        let au_count = header_bits / au_header_bits;
        let headers = payload
            .get(2..2 + header_bytes)
            .ok_or(DepacketizeError::Truncated)?;
        let mut data_off = 2 + header_bytes;
        let mut out = Vec::with_capacity(au_count);
        for i in 0..au_count {
            // Each AU-header is `au_header_bits` wide; for AAC-hbr that is 16
            // bits, so the size is the top `size_length` bits of a 2-byte field.
            let bit = i * au_header_bits;
            let byte = bit / 8;
            let hdr = headers
                .get(byte..byte + 2)
                .ok_or(DepacketizeError::Truncated)?;
            let size = (u16::from_be_bytes([hdr[0], hdr[1]]) >> (16 - self.size_length)) as usize;
            let end = data_off + size;
            let au = payload
                .get(data_off..end)
                .ok_or(DepacketizeError::Truncated)?;
            out.push(Bytes::copy_from_slice(au));
            data_off = end;
        }
        Ok(out)
    }
}

/// Packetizes H.264 Annex-B access units into RFC 6184 RTP packets — the inverse
/// of [`H264Depacketizer`], used for WebRTC/WHEP egress.
///
/// Each NAL unit that fits the MTU is sent as a single-NAL packet; larger NALs
/// are split into FU-A fragments. The RTP marker bit is set on the last packet
/// of each access unit so the receiver knows the frame is complete.
#[derive(Debug, Clone)]
pub struct RtpPacketizer {
    payload_type: u8,
    ssrc: u32,
    sequence: u16,
    /// Maximum RTP payload size (excluding the 12-byte header).
    max_payload: usize,
    /// Which NAL-based RTP payload format to emit (H.264 RFC 6184 vs H.265
    /// RFC 7798 — they differ in NAL-header width and FU framing).
    codec: NalCodec,
}

/// The NAL-based codecs [`RtpPacketizer`] / the depacketizers understand. Both
/// are Annex-B start-code framed; they differ in NAL-header width (1 vs 2 bytes)
/// and fragmentation-unit layout.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum NalCodec {
    H264,
    H265,
}

impl RtpPacketizer {
    /// An H.264 packetizer for `payload_type`/`ssrc`. `mtu` is the maximum UDP
    /// payload (1200 is the WebRTC-safe default); the 12-byte RTP header is
    /// subtracted.
    pub fn new(payload_type: u8, ssrc: u32, mtu: usize) -> Self {
        Self::with_codec(payload_type, ssrc, mtu, NalCodec::H264)
    }

    /// An H.265 (HEVC) packetizer, emitting the RFC 7798 payload format
    /// (2-byte NAL header, FU type 49). Counterpart to [`new`](Self::new).
    pub fn new_h265(payload_type: u8, ssrc: u32, mtu: usize) -> Self {
        Self::with_codec(payload_type, ssrc, mtu, NalCodec::H265)
    }

    fn with_codec(payload_type: u8, ssrc: u32, mtu: usize, codec: NalCodec) -> Self {
        Self {
            payload_type,
            ssrc,
            sequence: 0,
            max_payload: mtu.saturating_sub(12).max(1),
            codec,
        }
    }

    /// Emit one RTP packet: recycle a buffer from `recycle` (reusing its heap
    /// allocation), write the 12-byte header, let `fill` append the payload, and
    /// push it onto `out`. Advances the sequence number.
    fn emit(
        &mut self,
        recycle: &mut Vec<Vec<u8>>,
        out: &mut Vec<Vec<u8>>,
        marker: bool,
        timestamp: u32,
        fill: impl FnOnce(&mut Vec<u8>),
    ) {
        let mut buf = recycle.pop().unwrap_or_default();
        buf.clear();
        write_rtp_header(
            &mut buf,
            self.payload_type,
            marker,
            self.sequence,
            timestamp,
            self.ssrc,
        );
        self.sequence = self.sequence.wrapping_add(1);
        fill(&mut buf);
        out.push(buf);
    }

    /// Packetize one Annex-B access unit at `timestamp` (90 kHz) into RTP packets.
    ///
    /// Each NAL that fits the MTU is sent as a single-NAL packet; larger NALs are
    /// fragmented (FU-A for H.264, FU for H.265). The marker bit is set on the
    /// last packet of the access unit.
    ///
    /// Allocating variant; prefer [`packetize_into`](Self::packetize_into) on the
    /// egress hot path to recycle packet buffers across frames.
    pub fn packetize(&mut self, access_unit: &[u8], timestamp: u32) -> Vec<Vec<u8>> {
        let mut out = Vec::new();
        self.packetize_into(access_unit, timestamp, &mut out);
        out
    }

    /// Packetize into a caller-owned buffer, recycling the `Vec<u8>` allocations
    /// it already holds from a previous frame.
    ///
    /// The hot-path contract: pass the *same* `out` every frame. On entry the
    /// previously-produced (already-sent) packet buffers are reclaimed as a free
    /// pool and refilled in place, so steady-state egress performs no per-packet
    /// heap allocation.
    pub fn packetize_into(&mut self, access_unit: &[u8], timestamp: u32, out: &mut Vec<Vec<u8>>) {
        // Reclaim last frame's buffers (now consumed by the caller) as a pool.
        let mut recycle = std::mem::take(out);
        // Both codecs are Annex-B start-code framed; the shared scanner yields the
        // NAL units (without start codes) in order.
        let nals: Vec<&[u8]> = crate::codec::nal::iter_nals(access_unit)
            .filter(|n| !n.is_empty())
            .collect();
        for (i, nal) in nals.iter().enumerate() {
            let last_nal = i + 1 == nals.len();
            if nal.len() <= self.max_payload {
                // Single NAL unit packet; marker on the final NAL of the AU.
                self.emit(&mut recycle, out, last_nal, timestamp, |b| {
                    b.extend_from_slice(nal)
                });
            } else {
                match self.codec {
                    NalCodec::H264 => {
                        self.fragment_fua(nal, timestamp, last_nal, &mut recycle, out)
                    }
                    NalCodec::H265 => {
                        self.fragment_fu_h265(nal, timestamp, last_nal, &mut recycle, out)
                    }
                }
            }
        }
    }

    /// Split one oversized NAL into FU-A fragments (RFC 6184 §5.8).
    fn fragment_fua(
        &mut self,
        nal: &[u8],
        timestamp: u32,
        last_nal: bool,
        recycle: &mut Vec<Vec<u8>>,
        out: &mut Vec<Vec<u8>>,
    ) {
        let nal_header = nal[0];
        let fu_indicator = (nal_header & 0xE0) | 28; // F|NRI from NAL, type 28
        let nal_type = nal_header & 0x1F;
        let body = &nal[1..];
        // Each fragment carries a 2-byte FU header (indicator + FU header).
        let chunk = self.max_payload.saturating_sub(2).max(1);
        let n_chunks = body.len().div_ceil(chunk);
        for (idx, part) in body.chunks(chunk).enumerate() {
            let start = idx == 0;
            let end = idx + 1 == n_chunks;
            let mut fu_header = nal_type;
            if start {
                fu_header |= 0x80;
            }
            if end {
                fu_header |= 0x40;
            }
            // Marker only on the very last fragment of the final NAL of the AU.
            self.emit(recycle, out, last_nal && end, timestamp, |pkt| {
                pkt.push(fu_indicator);
                pkt.push(fu_header);
                pkt.extend_from_slice(part);
            });
        }
    }

    /// Split one oversized H.265 NAL into FU fragments (RFC 7798 §4.4.3).
    ///
    /// The 2-byte NAL header becomes a PayloadHdr with its type field set to 49
    /// (the F/LayerId/TID bits are preserved); each fragment then carries a
    /// 1-byte FU header (`S | E | FuType`) where `FuType` is the original type.
    fn fragment_fu_h265(
        &mut self,
        nal: &[u8],
        timestamp: u32,
        last_nal: bool,
        recycle: &mut Vec<Vec<u8>>,
        out: &mut Vec<Vec<u8>>,
    ) {
        // A well-formed H.265 NAL has a 2-byte header; anything shorter can't be
        // fragmented meaningfully, so emit it as a single packet.
        if nal.len() < 2 {
            self.emit(recycle, out, last_nal, timestamp, |pkt| {
                pkt.extend_from_slice(nal)
            });
            return;
        }
        let nal_type = (nal[0] >> 1) & 0x3F;
        // PayloadHdr: keep F (bit 15) + LayerId + TID, overwrite the 6-bit type
        // field with 49 (FU). Type occupies bits 9..14, i.e. (byte0 >> 1) & 0x3F.
        let payload_hdr0 = (nal[0] & 0x81) | (49 << 1);
        let payload_hdr1 = nal[1];
        let body = &nal[2..];
        // Each fragment carries the 2-byte PayloadHdr + 1-byte FU header.
        let chunk = self.max_payload.saturating_sub(3).max(1);
        let n_chunks = body.len().div_ceil(chunk);
        for (idx, part) in body.chunks(chunk).enumerate() {
            let start = idx == 0;
            let end = idx + 1 == n_chunks;
            let mut fu_header = nal_type;
            if start {
                fu_header |= 0x80;
            }
            if end {
                fu_header |= 0x40;
            }
            self.emit(recycle, out, last_nal && end, timestamp, |pkt| {
                pkt.push(payload_hdr0);
                pkt.push(payload_hdr1);
                pkt.push(fu_header);
                pkt.extend_from_slice(part);
            });
        }
    }
}

/// Reassembles RFC 6184 H.264 RTP payloads into Annex-B access units.
///
/// Feed each packet's payload (the bytes after [`RtpHeader::payload_offset`])
/// with its marker bit and timestamp to [`push`](Self::push). When a complete
/// access unit is ready the method returns `Ok(Some(au))`, where `au` is the
/// concatenated NAL units each prefixed with a 4-byte Annex-B start code —
/// exactly the shape the codec parsers and `annexb_to_avcc` expect.
#[derive(Debug, Default)]
pub struct H264Depacketizer {
    /// Bytes accumulated for the current access unit (Annex-B framed).
    au: Vec<u8>,
    /// FU-A reassembly buffer for the NAL currently being defragmented.
    fua: Vec<u8>,
    /// `true` while an FU-A fragment is in progress (between Start and End bits).
    in_fragment: bool,
    /// Reconstructed NAL header byte for the in-progress FU-A NAL.
    fua_header: u8,
    /// Timestamp of the access unit currently being assembled.
    current_ts: Option<u32>,
    /// Last sequence number seen (for gap detection during fragmentation).
    last_seq: Option<u16>,
}

impl H264Depacketizer {
    /// A fresh depacketizer with no in-progress access unit.
    pub fn new() -> Self {
        Self::default()
    }

    /// Append one NAL unit (Annex-B framed) to the current access unit.
    fn append_nal(&mut self, nal: &[u8]) {
        self.au.extend_from_slice(&ANNEXB_START);
        self.au.extend_from_slice(nal);
    }

    /// Whether the pending access unit holds an IDR (type 5) NAL — a keyframe.
    fn pending_is_keyframe(&self) -> bool {
        // Scan the assembled Annex-B for a NAL header with type 5.
        let mut i = 0;
        while i + 4 < self.au.len() {
            if self.au[i..i + 4] == ANNEXB_START {
                let nal_type = self.au[i + 4] & 0x1F;
                if nal_type == 5 {
                    return true;
                }
            }
            i += 1;
        }
        false
    }

    /// Emit and reset the pending access unit, if any.
    fn take_au(&mut self) -> Option<AccessUnit> {
        if self.au.is_empty() {
            return None;
        }
        let keyframe = self.pending_is_keyframe();
        let timestamp = self.current_ts.unwrap_or(0);
        let data = Bytes::from(std::mem::take(&mut self.au));
        self.current_ts = None;
        Some(AccessUnit {
            data,
            timestamp,
            keyframe,
        })
    }

    /// Push one RTP H.264 payload. Returns a completed [`AccessUnit`] when the
    /// marker bit closes the frame (or the timestamp advances to a new one).
    pub fn push(
        &mut self,
        payload: &[u8],
        marker: bool,
        timestamp: u32,
        sequence: u16,
    ) -> Result<Option<AccessUnit>, DepacketizeError> {
        if payload.is_empty() {
            return Err(DepacketizeError::Truncated);
        }

        // A timestamp change flushes the previous access unit before starting the
        // new one (some encoders omit the marker bit).
        let mut completed = None;
        if let Some(ts) = self.current_ts {
            if ts != timestamp && !self.in_fragment {
                completed = self.take_au();
            }
        }
        self.current_ts = Some(timestamp);

        let nal_type = payload[0] & 0x1F;
        match nal_type {
            1..=23 => {
                // Single NAL unit packet — the payload *is* the NAL.
                self.append_nal(payload);
            }
            24 => {
                // STAP-A: one byte type, then [u16 size][nal]… aggregates.
                let mut i = 1;
                while i + 2 <= payload.len() {
                    let size = u16::from_be_bytes([payload[i], payload[i + 1]]) as usize;
                    i += 2;
                    if i + size > payload.len() {
                        return Err(DepacketizeError::Truncated);
                    }
                    self.append_nal(&payload[i..i + size]);
                    i += size;
                }
            }
            28 => {
                // FU-A: byte0 = FU indicator, byte1 = FU header (S|E|R|type).
                if payload.len() < 2 {
                    return Err(DepacketizeError::Truncated);
                }
                let fu_header = payload[1];
                let start = fu_header & 0x80 != 0;
                let end = fu_header & 0x40 != 0;
                let frag_type = fu_header & 0x1F;

                if start {
                    // Reconstruct the original NAL header: F|NRI from the indicator,
                    // type from the FU header.
                    self.fua_header = (payload[0] & 0xE0) | frag_type;
                    self.fua.clear();
                    self.fua.push(self.fua_header);
                    self.in_fragment = true;
                } else if !self.in_fragment {
                    // Mid/last fragment with no start — lost the head.
                    return Err(DepacketizeError::OutOfOrder);
                } else if self.seq_gap(sequence) {
                    self.in_fragment = false;
                    self.fua.clear();
                    return Err(DepacketizeError::OutOfOrder);
                }
                self.fua.extend_from_slice(&payload[2..]);

                if end && self.in_fragment {
                    let nal = std::mem::take(&mut self.fua);
                    self.append_nal(&nal);
                    self.in_fragment = false;
                }
            }
            other => return Err(DepacketizeError::Unsupported(other)),
        }

        self.last_seq = Some(sequence);

        if completed.is_some() {
            return Ok(completed);
        }
        if marker {
            return Ok(self.take_au());
        }
        Ok(None)
    }

    /// Detect a one-step sequence-number gap relative to the previous packet.
    fn seq_gap(&self, sequence: u16) -> bool {
        match self.last_seq {
            Some(prev) => sequence.wrapping_sub(prev) != 1,
            None => false,
        }
    }
}

/// Reassembles RFC 7798 H.265 (HEVC) RTP payloads into Annex-B access units.
///
/// The H.265 counterpart to [`H264Depacketizer`], used for ingesting HEVC IP
/// cameras and encoders over RTSP/WebRTC. It handles the three packetization
/// modes: single NAL units, aggregation packets (AP, type 48), and fragmentation
/// units (FU, type 49). The output shape — NAL units each prefixed with a 4-byte
/// Annex-B start code — matches [`H264Depacketizer`] and the codec parsers.
///
/// DONL/DOND fields (only present when `sprop-max-don-diff > 0` is negotiated in
/// SDP) are not consumed; the common single-stream profile does not use them.
#[derive(Debug, Default)]
pub struct H265Depacketizer {
    /// Bytes accumulated for the current access unit (Annex-B framed).
    au: Vec<u8>,
    /// FU reassembly buffer for the NAL currently being defragmented.
    fu: Vec<u8>,
    /// `true` while an FU is in progress (between Start and End bits).
    in_fragment: bool,
    /// Timestamp of the access unit currently being assembled.
    current_ts: Option<u32>,
    /// Last sequence number seen (for gap detection during fragmentation).
    last_seq: Option<u16>,
}

impl H265Depacketizer {
    /// A fresh depacketizer with no in-progress access unit.
    pub fn new() -> Self {
        Self::default()
    }

    /// Append one NAL unit (Annex-B framed) to the current access unit.
    fn append_nal(&mut self, nal: &[u8]) {
        self.au.extend_from_slice(&ANNEXB_START);
        self.au.extend_from_slice(nal);
    }

    /// Whether the pending access unit holds an IRAP (BLA/IDR/CRA, types 16–23)
    /// VCL NAL — i.e. a random-access point / keyframe.
    fn pending_is_keyframe(&self) -> bool {
        let mut i = 0;
        while i + 4 < self.au.len() {
            if self.au[i..i + 4] == ANNEXB_START {
                let nal_type = (self.au[i + 4] >> 1) & 0x3F;
                if (16..=23).contains(&nal_type) {
                    return true;
                }
            }
            i += 1;
        }
        false
    }

    /// Emit and reset the pending access unit, if any.
    fn take_au(&mut self) -> Option<AccessUnit> {
        if self.au.is_empty() {
            return None;
        }
        let keyframe = self.pending_is_keyframe();
        let timestamp = self.current_ts.unwrap_or(0);
        let data = Bytes::from(std::mem::take(&mut self.au));
        self.current_ts = None;
        Some(AccessUnit {
            data,
            timestamp,
            keyframe,
        })
    }

    /// Detect a one-step sequence-number gap relative to the previous packet.
    fn seq_gap(&self, sequence: u16) -> bool {
        match self.last_seq {
            Some(prev) => sequence.wrapping_sub(prev) != 1,
            None => false,
        }
    }

    /// Push one RTP H.265 payload. Returns a completed [`AccessUnit`] when the
    /// marker bit closes the frame (or the timestamp advances to a new one).
    pub fn push(
        &mut self,
        payload: &[u8],
        marker: bool,
        timestamp: u32,
        sequence: u16,
    ) -> Result<Option<AccessUnit>, DepacketizeError> {
        // The H.265 NAL header is two bytes; a single byte cannot carry a type.
        if payload.len() < 2 {
            return Err(DepacketizeError::Truncated);
        }

        // A timestamp change flushes the previous access unit (some encoders omit
        // the marker bit).
        let mut completed = None;
        if let Some(ts) = self.current_ts {
            if ts != timestamp && !self.in_fragment {
                completed = self.take_au();
            }
        }
        self.current_ts = Some(timestamp);

        let nal_type = (payload[0] >> 1) & 0x3F;
        match nal_type {
            // Single NAL unit packet — the payload *is* the NAL (header included).
            0..=47 => self.append_nal(payload),
            48 => {
                // AP: 2-byte header, then [u16 size][nal]… aggregates.
                let mut i = 2;
                while i + 2 <= payload.len() {
                    let size = u16::from_be_bytes([payload[i], payload[i + 1]]) as usize;
                    i += 2;
                    if i + size > payload.len() {
                        return Err(DepacketizeError::Truncated);
                    }
                    self.append_nal(&payload[i..i + size]);
                    i += size;
                }
            }
            49 => {
                // FU: 2-byte PayloadHdr, 1-byte FU header (S|E|FuType), then body.
                if payload.len() < 3 {
                    return Err(DepacketizeError::Truncated);
                }
                let fu_header = payload[2];
                let start = fu_header & 0x80 != 0;
                let end = fu_header & 0x40 != 0;
                let fu_type = fu_header & 0x3F;

                if start {
                    // Reconstruct the original 2-byte NAL header: restore the type
                    // field (bits 9..14) from FuType, keep F/LayerId/TID.
                    let hdr0 = (payload[0] & 0x81) | (fu_type << 1);
                    let hdr1 = payload[1];
                    self.fu.clear();
                    self.fu.push(hdr0);
                    self.fu.push(hdr1);
                    self.in_fragment = true;
                } else if !self.in_fragment {
                    return Err(DepacketizeError::OutOfOrder);
                } else if self.seq_gap(sequence) {
                    self.in_fragment = false;
                    self.fu.clear();
                    return Err(DepacketizeError::OutOfOrder);
                }
                self.fu.extend_from_slice(&payload[3..]);

                if end && self.in_fragment {
                    let nal = std::mem::take(&mut self.fu);
                    self.append_nal(&nal);
                    self.in_fragment = false;
                }
            }
            other => return Err(DepacketizeError::Unsupported(other)),
        }

        self.last_seq = Some(sequence);

        if completed.is_some() {
            return Ok(completed);
        }
        if marker {
            return Ok(self.take_au());
        }
        Ok(None)
    }
}

/// A reassembled coded video frame from the RTP bus.
///
/// For the NAL codecs (H.264/H.265) `data` is the access unit in Annex-B form
/// (each NAL prefixed with a 4-byte start code); for VP9/AV1 it is the raw coded
/// frame / temporal unit. `keyframe` marks a decodable random-access point.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct AccessUnit {
    /// The coded frame bytes (Annex-B NALs for H.26x; raw frame for VP9/AV1).
    pub data: Bytes,
    /// RTP media timestamp (90 kHz) of the frame.
    pub timestamp: u32,
    /// Whether this is a keyframe / random-access point.
    pub keyframe: bool,
}

/// Write the 12-byte RTP fixed header (V=2, no padding/extension/CSRC) for one
/// packet. Shared by every packetizer in this module.
fn write_rtp_header(out: &mut Vec<u8>, pt: u8, marker: bool, seq: u16, ts: u32, ssrc: u32) {
    out.push(0x80); // V=2, P=0, X=0, CC=0
    out.push(if marker { 0x80 } else { 0 } | (pt & 0x7F));
    out.extend_from_slice(&seq.to_be_bytes());
    out.extend_from_slice(&ts.to_be_bytes());
    out.extend_from_slice(&ssrc.to_be_bytes());
}

// ── Opus (RFC 7587) ──────────────────────────────────────────────────────────

/// Packetizes Opus audio into RTP: each Opus packet is carried verbatim as the
/// payload of one RTP packet (RFC 7587 — Opus is self-delimiting, so there is no
/// payload descriptor). The 48 kHz media clock means the caller passes a
/// timestamp already scaled to 48 kHz.
#[derive(Debug, Clone)]
pub struct OpusPacketizer {
    payload_type: u8,
    ssrc: u32,
    sequence: u16,
}

impl OpusPacketizer {
    /// An Opus packetizer for `payload_type`/`ssrc`.
    pub fn new(payload_type: u8, ssrc: u32) -> Self {
        Self {
            payload_type,
            ssrc,
            sequence: 0,
        }
    }

    /// Packetize one Opus frame at `timestamp` (48 kHz) into a single RTP packet,
    /// appended to `out` (recycling a buffer it already holds). The RTP marker is
    /// left clear — continuous audio is not a talkspurt boundary.
    pub fn packetize_into(&mut self, opus: &[u8], timestamp: u32, out: &mut Vec<Vec<u8>>) {
        let mut recycle = std::mem::take(out);
        let mut pkt = recycle.pop().unwrap_or_default();
        pkt.clear();
        write_rtp_header(
            &mut pkt,
            self.payload_type,
            false,
            self.sequence,
            timestamp,
            self.ssrc,
        );
        self.sequence = self.sequence.wrapping_add(1);
        pkt.extend_from_slice(opus);
        out.push(pkt);
    }
}

// ── VP9 (draft-ietf-payload-vp9) ─────────────────────────────────────────────

/// Packetizes VP9 coded frames into RTP, using a flexible-mode-off payload
/// descriptor for the common single-layer (non-scalable) case.
///
/// Each frame is carried verbatim after a VP9 payload descriptor (the bytes are
/// not transformed — VP9 RTP carries the frame opaquely), split across the MTU
/// with the B (begin) bit on the first packet and E (end) + RTP marker on the
/// last. A 15-bit picture ID increments per frame. Spatial/temporal scalability
/// and flexible mode are out of scope.
#[derive(Debug, Clone)]
pub struct Vp9Packetizer {
    payload_type: u8,
    ssrc: u32,
    sequence: u16,
    max_payload: usize,
    picture_id: u16,
}

impl Vp9Packetizer {
    /// A VP9 packetizer for `payload_type`/`ssrc`. `mtu` is the maximum UDP
    /// payload; the 12-byte RTP header and a 3-byte descriptor are subtracted.
    pub fn new(payload_type: u8, ssrc: u32, mtu: usize) -> Self {
        Self {
            payload_type,
            ssrc,
            sequence: 0,
            // 12-byte RTP header + up to 3-byte descriptor (1 flags + 2 picture id).
            max_payload: mtu.saturating_sub(12 + 3).max(1),
            picture_id: 0,
        }
    }

    /// Packetize one VP9 frame at `timestamp` (90 kHz). `keyframe` clears the P
    /// (inter-predicted) bit so receivers can identify random-access points.
    pub fn packetize(&mut self, frame: &[u8], timestamp: u32, keyframe: bool) -> Vec<Vec<u8>> {
        let mut out = Vec::new();
        self.packetize_into(frame, timestamp, keyframe, &mut out);
        out
    }

    /// Recycling variant of [`packetize`](Self::packetize): pass the same `out`
    /// every frame to reuse the packet-buffer allocations across frames.
    pub fn packetize_into(
        &mut self,
        frame: &[u8],
        timestamp: u32,
        keyframe: bool,
        out: &mut Vec<Vec<u8>>,
    ) {
        let pid = self.picture_id & 0x7FFF;
        self.picture_id = self.picture_id.wrapping_add(1);

        let mut recycle = std::mem::take(out);
        let chunks: Vec<&[u8]> = if frame.is_empty() {
            vec![&[]]
        } else {
            frame.chunks(self.max_payload).collect()
        };
        let n = chunks.len();
        for (i, chunk) in chunks.into_iter().enumerate() {
            let begin = i == 0;
            let end = i + 1 == n;
            let mut pkt = recycle.pop().unwrap_or_default();
            pkt.clear();
            write_rtp_header(
                &mut pkt,
                self.payload_type,
                end,
                self.sequence,
                timestamp,
                self.ssrc,
            );
            self.sequence = self.sequence.wrapping_add(1);

            // Descriptor octet: I=1, P=!keyframe, L=0, F=0, B, E, V=0, Z=0.
            let mut desc0 = 0x80; // I = 1 (picture ID present)
            if !keyframe {
                desc0 |= 0x40; // P (inter-predicted)
            }
            if begin {
                desc0 |= 0x08; // B (start of frame)
            }
            if end {
                desc0 |= 0x04; // E (end of frame)
            }
            pkt.push(desc0);
            // 15-bit picture ID (M=1): 0x80|hi, lo.
            pkt.push(0x80 | (pid >> 8) as u8);
            pkt.push((pid & 0xFF) as u8);
            pkt.extend_from_slice(chunk);
            out.push(pkt);
        }
    }
}

/// Reassembles VP9 RTP payloads (draft-ietf-payload-vp9, non-flexible single
/// layer) into coded frames. Counterpart to [`Vp9Packetizer`].
#[derive(Debug, Default)]
pub struct Vp9Depacketizer {
    frame: Vec<u8>,
    in_frame: bool,
    keyframe: bool,
    current_ts: Option<u32>,
}

impl Vp9Depacketizer {
    /// A fresh depacketizer with no in-progress frame.
    pub fn new() -> Self {
        Self::default()
    }

    /// Push one VP9 RTP payload. Returns a completed frame when the E (end) bit
    /// and RTP marker close it.
    pub fn push(
        &mut self,
        payload: &[u8],
        marker: bool,
        timestamp: u32,
    ) -> Result<Option<AccessUnit>, DepacketizeError> {
        if payload.is_empty() {
            return Err(DepacketizeError::Truncated);
        }
        let desc0 = payload[0];
        let has_pid = desc0 & 0x80 != 0;
        let has_layer = desc0 & 0x20 != 0;
        let flexible = desc0 & 0x10 != 0;
        let begin = desc0 & 0x08 != 0;
        let end = desc0 & 0x04 != 0;
        let predicted = desc0 & 0x40 != 0;

        // Walk past the variable-length descriptor fields we recognize.
        let mut off = 1;
        if has_pid {
            // M bit selects a 1- or 2-byte picture ID.
            let m = payload.get(off).ok_or(DepacketizeError::Truncated)? & 0x80 != 0;
            off += if m { 2 } else { 1 };
        }
        if has_layer {
            off += 1; // TID/U/SID/D byte
            if !flexible {
                off += 1; // TL0PICIDX (non-flexible mode)
            }
        }
        if off > payload.len() {
            return Err(DepacketizeError::Truncated);
        }

        if begin {
            self.frame.clear();
            self.in_frame = true;
            self.keyframe = !predicted;
            self.current_ts = Some(timestamp);
        } else if !self.in_frame {
            return Err(DepacketizeError::OutOfOrder);
        }
        self.frame.extend_from_slice(&payload[off..]);

        if end && marker && self.in_frame {
            self.in_frame = false;
            return Ok(Some(AccessUnit {
                data: Bytes::from(std::mem::take(&mut self.frame)),
                timestamp: self.current_ts.unwrap_or(timestamp),
                keyframe: self.keyframe,
            }));
        }
        Ok(None)
    }
}

// ── AV1 (AOMedia "RTP Payload Format For AV1") ───────────────────────────────

/// Encode `v` as unsigned LEB128 into `out`.
#[cfg(feature = "codec-av1")]
fn leb128_encode(mut v: u64, out: &mut Vec<u8>) {
    loop {
        let mut byte = (v & 0x7F) as u8;
        v >>= 7;
        if v != 0 {
            byte |= 0x80;
        }
        out.push(byte);
        if v == 0 {
            break;
        }
    }
}

#[cfg(feature = "codec-av1")]
const AV1_OBU_SEQUENCE_HEADER: u8 = 1;
#[cfg(feature = "codec-av1")]
const AV1_OBU_TEMPORAL_DELIMITER: u8 = 2;

/// Packetizes an AV1 temporal unit into RTP using the AOMedia payload format.
///
/// Each non-temporal-delimiter OBU is re-framed as a length-delimited *OBU
/// element* (the W=0 form, with the OBU's `obu_has_size_field` cleared), the
/// elements are concatenated, and the resulting stream is split across the MTU
/// with a one-byte aggregation header per packet (Z/Y continuation bits, N on a
/// new coded video sequence). The temporal delimiter is dropped per the spec;
/// frame boundaries are conveyed by the RTP marker. Scalability structures are
/// not emitted.
#[cfg(feature = "codec-av1")]
#[derive(Debug, Clone)]
pub struct Av1Packetizer {
    payload_type: u8,
    ssrc: u32,
    sequence: u16,
    max_payload: usize,
}

#[cfg(feature = "codec-av1")]
impl Av1Packetizer {
    /// An AV1 packetizer for `payload_type`/`ssrc`. `mtu` is the maximum UDP
    /// payload; the 12-byte RTP header and 1-byte aggregation header are removed.
    pub fn new(payload_type: u8, ssrc: u32, mtu: usize) -> Self {
        Self {
            payload_type,
            ssrc,
            sequence: 0,
            max_payload: mtu.saturating_sub(12 + 1).max(1),
        }
    }

    /// Packetize one AV1 temporal unit (low-overhead OBUs) at `timestamp`.
    pub fn packetize(&mut self, temporal_unit: &[u8], timestamp: u32) -> Vec<Vec<u8>> {
        let mut out = Vec::new();
        self.packetize_into(temporal_unit, timestamp, &mut out);
        out
    }

    /// Recycling variant of [`packetize`](Self::packetize): pass the same `out`
    /// every temporal unit to reuse the packet-buffer allocations.
    pub fn packetize_into(&mut self, temporal_unit: &[u8], timestamp: u32, out: &mut Vec<Vec<u8>>) {
        // Re-frame each OBU (minus the temporal delimiter) as a length-delimited
        // OBU element with obu_has_size_field cleared.
        let mut stream = Vec::with_capacity(temporal_unit.len());
        let mut new_cvs = false;
        for obu in crate::codec::obu::iter_obus(temporal_unit) {
            if obu.obu_type == AV1_OBU_TEMPORAL_DELIMITER {
                continue;
            }
            if obu.obu_type == AV1_OBU_SEQUENCE_HEADER {
                new_cvs = true;
            }
            let header_len = 1 + obu.has_extension as usize;
            let mut element = Vec::with_capacity(header_len + obu.payload.len());
            element.push(obu.raw[0] & !0x02); // clear obu_has_size_field
            if obu.has_extension {
                element.push(obu.raw[1]);
            }
            element.extend_from_slice(obu.payload);
            leb128_encode(element.len() as u64, &mut stream);
            stream.extend_from_slice(&element);
        }

        let mut recycle = std::mem::take(out);
        let chunks: Vec<&[u8]> = if stream.is_empty() {
            vec![&[]]
        } else {
            stream.chunks(self.max_payload).collect()
        };
        let n = chunks.len();
        for (i, chunk) in chunks.into_iter().enumerate() {
            let last = i + 1 == n;
            let mut pkt = recycle.pop().unwrap_or_default();
            pkt.clear();
            write_rtp_header(
                &mut pkt,
                self.payload_type,
                last,
                self.sequence,
                timestamp,
                self.ssrc,
            );
            self.sequence = self.sequence.wrapping_add(1);

            // Aggregation header: Z (continues previous packet) | Y (continues in
            // next) | W=0 (length-delimited elements) | N (new coded video seq).
            let mut agg = 0u8;
            if i > 0 {
                agg |= 0x80; // Z
            }
            if !last {
                agg |= 0x40; // Y
            }
            if i == 0 && new_cvs {
                agg |= 0x08; // N
            }
            pkt.push(agg);
            pkt.extend_from_slice(chunk);
            out.push(pkt);
        }
    }
}

/// Reassembles AV1 RTP payloads into temporal units. Counterpart to
/// [`Av1Packetizer`]: it concatenates each packet's OBU-element bytes (past the
/// aggregation header) and, on the RTP marker, parses the length-delimited
/// elements back into low-overhead OBUs (re-adding each `obu_has_size_field`).
#[cfg(feature = "codec-av1")]
#[derive(Debug, Default)]
pub struct Av1Depacketizer {
    stream: Vec<u8>,
    new_cvs: bool,
    current_ts: Option<u32>,
}

#[cfg(feature = "codec-av1")]
impl Av1Depacketizer {
    /// A fresh depacketizer with no in-progress temporal unit.
    pub fn new() -> Self {
        Self::default()
    }

    /// Push one AV1 RTP payload. Returns a completed temporal unit when the RTP
    /// marker closes it.
    pub fn push(
        &mut self,
        payload: &[u8],
        marker: bool,
        timestamp: u32,
    ) -> Result<Option<AccessUnit>, DepacketizeError> {
        if payload.is_empty() {
            return Err(DepacketizeError::Truncated);
        }
        let agg = payload[0];
        if agg & 0x08 != 0 {
            self.new_cvs = true; // N: new coded video sequence
        }
        if self.current_ts.is_none() {
            self.current_ts = Some(timestamp);
        }
        self.stream.extend_from_slice(&payload[1..]);

        if !marker {
            return Ok(None);
        }

        // Marker: rebuild the temporal unit from length-delimited OBU elements.
        let stream = std::mem::take(&mut self.stream);
        let mut tu = Vec::with_capacity(stream.len() + 8);
        let mut pos = 0;
        while pos < stream.len() {
            let len = leb128_decode(&stream, &mut pos).ok_or(DepacketizeError::Truncated)?;
            let end = pos.checked_add(len).ok_or(DepacketizeError::Truncated)?;
            let element = stream.get(pos..end).ok_or(DepacketizeError::Truncated)?;
            pos = end;
            // Element → low-overhead OBU: set obu_has_size_field, insert the size.
            let hdr0 = *element.first().ok_or(DepacketizeError::Truncated)?;
            let has_ext = (hdr0 >> 2) & 1 == 1;
            let header_len = 1 + has_ext as usize;
            let obu_payload = element
                .get(header_len..)
                .ok_or(DepacketizeError::Truncated)?;
            tu.push(hdr0 | 0x02);
            if has_ext {
                tu.push(element[1]);
            }
            leb128_encode(obu_payload.len() as u64, &mut tu);
            tu.extend_from_slice(obu_payload);
        }

        let keyframe = std::mem::take(&mut self.new_cvs);
        let ts = self.current_ts.take().unwrap_or(timestamp);
        Ok(Some(AccessUnit {
            data: Bytes::from(tu),
            timestamp: ts,
            keyframe,
        }))
    }
}

/// Decode an unsigned LEB128 integer at `*pos` (advancing it), returning `None`
/// on truncation or overflow. Mirrors the codec-side decoder for RTP carriage.
#[cfg(feature = "codec-av1")]
fn leb128_decode(data: &[u8], pos: &mut usize) -> Option<usize> {
    let mut value: u64 = 0;
    for i in 0..8 {
        let byte = *data.get(*pos)?;
        *pos += 1;
        value |= ((byte & 0x7F) as u64) << (i * 7);
        if byte & 0x80 == 0 {
            return usize::try_from(value).ok();
        }
    }
    None
}

#[cfg(test)]
mod tests {
    use super::*;

    /// Build a minimal 12-byte RTP packet with the given fields and payload.
    fn rtp(seq: u16, ts: u32, marker: bool, payload: &[u8]) -> Vec<u8> {
        let mut p = vec![0x80, if marker { 0x80 | 96 } else { 96 }];
        p.extend_from_slice(&seq.to_be_bytes());
        p.extend_from_slice(&ts.to_be_bytes());
        p.extend_from_slice(&[0, 0, 0, 1]); // ssrc
        p.extend_from_slice(payload);
        p
    }

    #[test]
    fn parses_fixed_header_and_payload_offset() {
        let pkt = rtp(7, 9000, true, &[0x65, 0xAA]);
        let h = RtpHeader::parse(&pkt).unwrap();
        assert_eq!(h.sequence, 7);
        assert_eq!(h.timestamp, 9000);
        assert!(h.marker);
        assert_eq!(h.payload_type, 96);
        assert_eq!(h.payload_offset, 12);
        assert_eq!(&pkt[h.payload_offset..], &[0x65, 0xAA]);
    }

    #[test]
    fn rejects_wrong_version_and_short_buffers() {
        assert!(RtpHeader::parse(&[0x00; 12]).is_none()); // version 0
        assert!(RtpHeader::parse(&[0x80; 4]).is_none()); // too short
    }

    #[test]
    fn honors_csrc_count_in_payload_offset() {
        let mut pkt = rtp(1, 0, false, &[0x41]);
        pkt[0] = 0x82; // version 2, CSRC count = 2
        let mut with_csrc = pkt[..12].to_vec();
        with_csrc.extend_from_slice(&[0xDE, 0xAD, 0xBE, 0xEF, 0, 0, 0, 0]); // 2 CSRCs
        with_csrc.push(0x41);
        let h = RtpHeader::parse(&with_csrc).unwrap();
        assert_eq!(h.payload_offset, 20);
    }

    #[test]
    fn aac_hbr_splits_two_access_units() {
        // AU-headers-length = 32 bits → two 16-bit AU-headers.
        // AU sizes 3 and 2 (top 13 bits of each 2-byte header).
        let mut p = Vec::new();
        p.extend_from_slice(&32u16.to_be_bytes()); // header bits
        p.extend_from_slice(&((3u16) << 3).to_be_bytes()); // AU-header: size 3
        p.extend_from_slice(&((2u16) << 3).to_be_bytes()); // AU-header: size 2
        p.extend_from_slice(&[0xA1, 0xA2, 0xA3]); // AU 1
        p.extend_from_slice(&[0xB1, 0xB2]); // AU 2
        let aus = AacDepacketizer::new().push(&p).unwrap();
        assert_eq!(aus.len(), 2);
        assert_eq!(&aus[0][..], &[0xA1, 0xA2, 0xA3]);
        assert_eq!(&aus[1][..], &[0xB1, 0xB2]);
    }

    #[test]
    fn aac_hbr_single_au() {
        let mut p = Vec::new();
        p.extend_from_slice(&16u16.to_be_bytes()); // one 16-bit AU-header
        p.extend_from_slice(&((4u16) << 3).to_be_bytes()); // size 4
        p.extend_from_slice(&[1, 2, 3, 4]);
        let aus = AacDepacketizer::new().push(&p).unwrap();
        assert_eq!(aus.len(), 1);
        assert_eq!(&aus[0][..], &[1, 2, 3, 4]);
    }

    #[test]
    fn aac_truncated_payload_errors() {
        assert_eq!(
            AacDepacketizer::new().push(&[0x00]),
            Err(DepacketizeError::Truncated)
        );
        // Declares one AU of size 8 but supplies only 2 data bytes.
        let mut p = 16u16.to_be_bytes().to_vec();
        p.extend_from_slice(&((8u16) << 3).to_be_bytes());
        p.extend_from_slice(&[1, 2]);
        assert_eq!(
            AacDepacketizer::new().push(&p),
            Err(DepacketizeError::Truncated)
        );
    }

    #[test]
    fn single_nal_packet_emits_annexb_on_marker() {
        let mut d = H264Depacketizer::new();
        // Type 1 (non-IDR slice), marker set → one access unit.
        let out = d.push(&[0x41, 0x9A, 0xBC], true, 3000, 1).unwrap().unwrap();
        assert_eq!(&out.data[..], &[0, 0, 0, 1, 0x41, 0x9A, 0xBC]);
        assert!(!out.keyframe);
        assert_eq!(out.timestamp, 3000);
    }

    #[test]
    fn idr_single_nal_is_flagged_keyframe() {
        let mut d = H264Depacketizer::new();
        let out = d.push(&[0x65, 0x01], true, 0, 1).unwrap().unwrap();
        assert!(out.keyframe);
    }

    #[test]
    fn packetizer_single_nal_round_trips_through_depacketizer() {
        // A small AU (two NALs) → single-NAL packets → reassembled identically.
        let au = [0, 0, 0, 1, 0x67, 0x42, 0x00, 0, 0, 0, 1, 0x65, 0x88, 0x99];
        let mut pkt = RtpPacketizer::new(96, 0xABCD, 1200);
        let packets = pkt.packetize(&au, 3000);
        assert_eq!(packets.len(), 2, "one packet per NAL");

        let mut depack = H264Depacketizer::new();
        let mut out = None;
        for p in &packets {
            let h = RtpHeader::parse(p).unwrap();
            if let Some(au) = depack
                .push(&p[h.payload_offset..], h.marker, h.timestamp, h.sequence)
                .unwrap()
            {
                out = Some(au);
            }
        }
        let out = out.expect("AU completed on the marker packet");
        assert_eq!(&out.data[..], &au);
        assert!(out.keyframe);
        assert_eq!(out.timestamp, 3000);
    }

    #[test]
    fn packetize_into_recycles_buffers_without_changing_output() {
        // The recycling hot-path API must produce byte-identical packets to the
        // allocating `packetize`, frame after frame, including correct sequence
        // numbers carried across the reused buffer.
        let au1 = [0, 0, 0, 1, 0x67, 0x42, 0x00, 0, 0, 0, 1, 0x65, 0x88, 0x99];
        let au2 = [0, 0, 0, 1, 0x65, 0x11, 0x22, 0x33];

        let mut a = RtpPacketizer::new(96, 0xABCD, 1200);
        let mut b = RtpPacketizer::new(96, 0xABCD, 1200);
        let mut reused: Vec<Vec<u8>> = Vec::new();

        for au in [&au1[..], &au2[..], &au1[..]] {
            let expected = a.packetize(au, 3000);
            // Capture the backing pointers to prove buffers are reused, not freed.
            b.packetize_into(au, 3000, &mut reused);
            assert_eq!(
                reused, expected,
                "recycled output matches allocating output"
            );
        }
    }

    #[test]
    fn packetizer_fragments_oversized_nal_and_round_trips() {
        // One NAL larger than the MTU → FU-A fragments → reassembled identically.
        let mut nal = vec![0, 0, 0, 1, 0x65]; // start code + IDR NAL header
        nal.extend((0..600u16).map(|i| i as u8)); // long payload
        let mut pkt = RtpPacketizer::new(96, 1, 100); // tiny MTU forces FU-A
        let packets = pkt.packetize(&nal, 90);
        assert!(packets.len() > 1, "oversized NAL is fragmented");
        // Only the last packet carries the marker bit.
        let markers: Vec<bool> = packets
            .iter()
            .map(|p| RtpHeader::parse(p).unwrap().marker)
            .collect();
        assert_eq!(markers.iter().filter(|m| **m).count(), 1);
        assert!(markers.last().unwrap());

        let mut depack = H264Depacketizer::new();
        let mut out = None;
        for p in &packets {
            let h = RtpHeader::parse(p).unwrap();
            if let Some(au) = depack
                .push(&p[h.payload_offset..], h.marker, h.timestamp, h.sequence)
                .unwrap()
            {
                out = Some(au);
            }
        }
        assert_eq!(&out.unwrap().data[..], &nal[..]);
    }

    #[test]
    fn stap_a_splits_aggregated_nals() {
        // STAP-A (24): [24][size=2][AA BB][size=3][CC DD EE]
        let payload = [24, 0, 2, 0xAA, 0xBB, 0, 3, 0xCC, 0xDD, 0xEE];
        let mut d = H264Depacketizer::new();
        let out = d.push(&payload, true, 0, 1).unwrap().unwrap();
        assert_eq!(
            &out.data[..],
            &[0, 0, 0, 1, 0xAA, 0xBB, 0, 0, 0, 1, 0xCC, 0xDD, 0xEE]
        );
    }

    #[test]
    fn fu_a_reassembles_fragmented_nal() {
        let mut d = H264Depacketizer::new();
        // FU indicator 0x7C (F=0,NRI=3,type=28), FU header start 0x85 (S=1,type=5).
        assert!(d
            .push(&[0x7C, 0x85, 0x11, 0x22], false, 0, 1)
            .unwrap()
            .is_none());
        // Middle fragment (S=0,E=0).
        assert!(d.push(&[0x7C, 0x05, 0x33], false, 0, 2).unwrap().is_none());
        // End fragment (E=1), marker closes the AU.
        let out = d.push(&[0x7C, 0x45, 0x44], true, 0, 3).unwrap().unwrap();
        // Reconstructed NAL header: NRI 0x60 | type 5 = 0x65, then payload bytes.
        assert_eq!(&out.data[..], &[0, 0, 0, 1, 0x65, 0x11, 0x22, 0x33, 0x44]);
        assert!(out.keyframe);
    }

    #[test]
    fn fu_a_sequence_gap_reports_out_of_order() {
        let mut d = H264Depacketizer::new();
        d.push(&[0x7C, 0x85, 0x11], false, 0, 1).unwrap();
        // Jump from seq 1 to seq 5 mid-fragment.
        assert_eq!(
            d.push(&[0x7C, 0x05, 0x22], false, 0, 5),
            Err(DepacketizeError::OutOfOrder)
        );
    }

    #[test]
    fn timestamp_change_flushes_previous_au_without_marker() {
        let mut d = H264Depacketizer::new();
        // First AU, no marker.
        assert!(d.push(&[0x41, 0x01], false, 1000, 1).unwrap().is_none());
        // New timestamp flushes the first AU.
        let out = d.push(&[0x41, 0x02], false, 2000, 2).unwrap().unwrap();
        assert_eq!(out.timestamp, 1000);
        assert_eq!(&out.data[..], &[0, 0, 0, 1, 0x41, 0x01]);
    }

    // ── H.265 (RFC 7798) ────────────────────────────────────────────────────
    // H.265 NAL headers are two bytes; type = (byte0 >> 1) & 0x3F. Examples used
    // below: VPS=32 (0x40,0x01), IDR_W_RADL=19 (0x26,0x01).

    #[test]
    fn h265_single_nal_round_trips_through_depacketizer() {
        // VPS (non-VCL) + IDR (VCL keyframe), each a single-NAL packet.
        let au = [
            0, 0, 0, 1, 0x40, 0x01, 0xAA, // VPS (type 32)
            0, 0, 0, 1, 0x26, 0x01, 0x88, 0x99, // IDR (type 19)
        ];
        let mut pkt = RtpPacketizer::new_h265(96, 0xABCD, 1200);
        let packets = pkt.packetize(&au, 3000);
        assert_eq!(packets.len(), 2, "one packet per NAL");

        let mut depack = H265Depacketizer::new();
        let mut out = None;
        for p in &packets {
            let h = RtpHeader::parse(p).unwrap();
            if let Some(au) = depack
                .push(&p[h.payload_offset..], h.marker, h.timestamp, h.sequence)
                .unwrap()
            {
                out = Some(au);
            }
        }
        let out = out.expect("AU completed on the marker packet");
        assert_eq!(&out.data[..], &au);
        assert!(out.keyframe, "IRAP type 19 is a keyframe");
        assert_eq!(out.timestamp, 3000);
    }

    #[test]
    fn h265_fragments_oversized_nal_and_round_trips() {
        // One IDR NAL larger than the MTU → FU fragments → reassembled identically.
        let mut nal = vec![0, 0, 0, 1, 0x26, 0x01]; // start code + 2-byte IDR header
        nal.extend((0..600u16).map(|i| i as u8));
        let mut pkt = RtpPacketizer::new_h265(96, 1, 100); // tiny MTU forces FU
        let packets = pkt.packetize(&nal, 90);
        assert!(packets.len() > 1, "oversized NAL is fragmented");
        // Exactly one marker, on the last fragment.
        let markers: Vec<bool> = packets
            .iter()
            .map(|p| RtpHeader::parse(p).unwrap().marker)
            .collect();
        assert_eq!(markers.iter().filter(|m| **m).count(), 1);
        assert!(markers.last().unwrap());
        // Each FU packet carries a type-49 PayloadHdr.
        for p in &packets {
            let h = RtpHeader::parse(p).unwrap();
            let pt = (p[h.payload_offset] >> 1) & 0x3F;
            assert_eq!(pt, 49, "FU payload type");
        }

        let mut depack = H265Depacketizer::new();
        let mut out = None;
        for p in &packets {
            let h = RtpHeader::parse(p).unwrap();
            if let Some(au) = depack
                .push(&p[h.payload_offset..], h.marker, h.timestamp, h.sequence)
                .unwrap()
            {
                out = Some(au);
            }
        }
        assert_eq!(&out.unwrap().data[..], &nal[..]);
    }

    #[test]
    fn h265_ap_splits_aggregated_nals() {
        // AP (type 48): [0x60,0x01][size=2][AA BB][size=3][CC DD EE]
        let payload = [0x60, 0x01, 0, 2, 0xAA, 0xBB, 0, 3, 0xCC, 0xDD, 0xEE];
        let mut d = H265Depacketizer::new();
        let out = d.push(&payload, true, 0, 1).unwrap().unwrap();
        assert_eq!(
            &out.data[..],
            &[0, 0, 0, 1, 0xAA, 0xBB, 0, 0, 0, 1, 0xCC, 0xDD, 0xEE]
        );
    }

    #[test]
    fn h265_rejects_truncated_and_unsupported() {
        let mut d = H265Depacketizer::new();
        // One byte cannot hold a 2-byte NAL header.
        assert_eq!(
            d.push(&[0x26], true, 0, 1),
            Err(DepacketizeError::Truncated)
        );
        // PACI (type 50) is not supported.
        assert_eq!(
            d.push(&[50 << 1, 0x01, 0x00], true, 0, 2),
            Err(DepacketizeError::Unsupported(50))
        );
    }

    // ── VP9 ───────────────────────────────────────────────────────────────────

    fn vp9_depacketize(packets: &[Vec<u8>]) -> Option<AccessUnit> {
        let mut d = Vp9Depacketizer::new();
        let mut out = None;
        for p in packets {
            let h = RtpHeader::parse(p).unwrap();
            if let Some(f) = d
                .push(&p[h.payload_offset..], h.marker, h.timestamp)
                .unwrap()
            {
                out = Some(f);
            }
        }
        out
    }

    #[test]
    fn vp9_fragmented_frame_round_trips() {
        let frame: Vec<u8> = (0..500u16).map(|i| i as u8).collect();
        let mut pkt = Vp9Packetizer::new(98, 0x1234, 100); // small MTU → fragments
        let packets = pkt.packetize(&frame, 9000, true);
        assert!(packets.len() > 1, "frame fragmented");

        // Exactly one marker, on the last packet.
        let markers: Vec<bool> = packets
            .iter()
            .map(|p| RtpHeader::parse(p).unwrap().marker)
            .collect();
        assert_eq!(markers.iter().filter(|m| **m).count(), 1);
        assert!(markers.last().unwrap());

        let out = vp9_depacketize(&packets).expect("frame completed");
        assert_eq!(&out.data[..], &frame[..]);
        assert!(out.keyframe, "keyframe → P bit clear");
        assert_eq!(out.timestamp, 9000);
    }

    #[test]
    fn vp9_inter_frame_is_not_a_keyframe() {
        let mut pkt = Vp9Packetizer::new(98, 1, 1200);
        let packets = pkt.packetize(&[1, 2, 3], 0, false);
        assert_eq!(packets.len(), 1);
        let out = vp9_depacketize(&packets).expect("frame");
        assert_eq!(&out.data[..], &[1, 2, 3]);
        assert!(!out.keyframe, "P bit set → inter frame");
    }

    // ── AV1 ───────────────────────────────────────────────────────────────────

    #[cfg(feature = "codec-av1")]
    fn av1_depacketize(packets: &[Vec<u8>]) -> Option<AccessUnit> {
        let mut d = Av1Depacketizer::new();
        let mut out = None;
        for p in packets {
            let h = RtpHeader::parse(p).unwrap();
            if let Some(f) = d
                .push(&p[h.payload_offset..], h.marker, h.timestamp)
                .unwrap()
            {
                out = Some(f);
            }
        }
        out
    }

    #[cfg(feature = "codec-av1")]
    #[test]
    fn av1_temporal_unit_round_trips_without_delimiter() {
        // Low-overhead OBUs: temporal delimiter + sequence header + frame.
        let td = [0x12u8, 0x00];
        let seq = [0x0Au8, 0x02, 0xAA, 0xBB];
        let frame = [0x32u8, 0x03, 0x11, 0x22, 0x33];
        let mut tu = Vec::new();
        tu.extend_from_slice(&td);
        tu.extend_from_slice(&seq);
        tu.extend_from_slice(&frame);

        let mut pkt = Av1Packetizer::new(99, 7, 1200);
        let packets = pkt.packetize(&tu, 1000);
        let out = av1_depacketize(&packets).expect("TU completed");

        // The temporal delimiter is dropped; seq + frame survive, low-overhead.
        let mut expected = Vec::new();
        expected.extend_from_slice(&seq);
        expected.extend_from_slice(&frame);
        assert_eq!(&out.data[..], &expected[..]);
        assert!(out.keyframe, "sequence header → new coded video sequence");
        assert_eq!(out.timestamp, 1000);
    }

    #[cfg(feature = "codec-av1")]
    #[test]
    fn av1_large_temporal_unit_fragments_and_round_trips() {
        // A frame OBU with a 300-byte payload (size field leb128(300) = AC 02).
        let mut frame = vec![0x32u8, 0xAC, 0x02];
        frame.extend((0..300u16).map(|i| i as u8));
        let mut tu = vec![0x12u8, 0x00]; // temporal delimiter
        tu.extend_from_slice(&frame);

        let mut pkt = Av1Packetizer::new(99, 1, 64); // tiny MTU forces fragmentation
        let packets = pkt.packetize(&tu, 0);
        assert!(packets.len() > 1, "large TU fragmented");
        // Z set on every packet but the first; Y on every packet but the last.
        for (i, p) in packets.iter().enumerate() {
            let agg = p[RtpHeader::parse(p).unwrap().payload_offset];
            assert_eq!((agg & 0x80 != 0), i > 0, "Z continuation bit");
            assert_eq!(
                (agg & 0x40 != 0),
                i + 1 < packets.len(),
                "Y continuation bit"
            );
        }

        let out = av1_depacketize(&packets).expect("TU completed");
        assert_eq!(&out.data[..], &frame[..], "frame OBU reconstructed");
    }
}