# VerseVDI complete media datagram v2

Datagram v2 carries one complete encoded video or audio unit under one sequence.
It is negotiated explicitly as `datagram-v2`; v1 bytes are never reinterpreted.
The gateway relays encoded bytes without codec processing or provider exposure.

All multi-byte integers are unsigned big-endian. The fixed header is 23 bytes:

| Offset | Size | Field | Rule |
|---:|---:|---|---|
| 0 | 2 | magic | ASCII `VD` (`0x56 0x44`) |
| 2 | 1 | version | `2` only |
| 3 | 1 | channel | `media.video.v1` (10) or `media.audio.v1` (11) |
| 4 | 1 | flags | zero; unknown bits reject |
| 5 | 4 | sequence | session-local wrapping complete-unit sequence |
| 9 | 8 | timestamp_ms | sender timestamp, bounded by transport skew policy |
| 17 | 2 | fragment_index | zero-based |
| 19 | 2 | fragment_count | 1 through 891; index less than count |
| 21 | 2 | payload_length | exact payload byte count, at most 1,177 |

Each QUIC datagram is at most 1,200 bytes. One complete unit is at most
1,048,576 encoded bytes and 891 fragments. A sender rejects a larger unit
before fragmentation.

A receiver retains at most four incomplete media units and only received
fragment bytes. Fragments for one unit must agree on channel, sequence,
timestamp, flags, and count. Exact duplicates are ignored; conflicting
duplicates reject that unit. Bounded reorder is accepted. An incomplete unit
expires after 250 milliseconds, and accepting a fifth incomplete unit evicts
the oldest. Reassembly checks the 1,048,576-byte ceiling before appending and
emits only after every fragment is present.