Args:
    source: URL of the input video (https://, s3://, or `tus:<uuid>`).
    outputs: list of format-spec dicts. Each MUST have an `output` field
        whose value is one of: mp4, webm, advanced_hls, advanced_dash,
        webm_dash, repack, mp3, m4a, hls_audio, flac, gif, thumbnail,
        thumbnails, smart_thumbnail, metadata, speech_to_text, vmaf,
        video_intelligence, ai_detection, waveform.
        The OUTER array is named `format` in the Qencode schema (this
        tool wraps it for you). The INNER STRING field naming the type
        is `output` — NOT `format`. This is the most common composition
        mistake. Example of a valid entry:
            {
                "output": "mp4",
                "video_codec": "libx264",
                "audio_codec": "libfdk_aac",
                "resolution": 720,
                "optimize_bitrate": 1,
                "audio_bitrate": 128,
                "destination": {
                    "url": "<destination_prefix>/<new filename or folder>"
                }
            }
        Media Storage `destination.url` is `{destination_prefix}/{new key}`
        — a NEW file or folder, never a listed object's `destination_url`
        (that is the existing file, for reading as `source`). Folder
        outputs (HLS / STT / VI / thumbnails) take `{destination_prefix}/folder`
        with no trailing slash. Do not assemble the host.
        `s3://<bucket>/…` is rejected here (the API would demand
        `key`/`secret`). Inputs the encoder fetches (`source`, `distorted`,
        subtitle files, `logo.source`) use `cdn_url` when public and
        `destination_url` when private. Video Intelligence `source` must
        be https — `cdn_url` only; a private object cannot be a VI source.
        For HLS/DASH ABR, put per-rendition params on each entry of an
        inner `stream[]` array (not on the format object directly).
        Output-specific required fields (see matching recipe):
            advanced_hls / advanced_dash / webm_dash / hls_audio —
                non-empty `stream[]` of objects. A bare
                `{"output": "advanced_hls"}` is rejected. Fetch
                `qencode://recipe/hls_abr` (or `audio_outputs` for
                `hls_audio`) before composing.
            vmaf — `distorted` URL of the encoded video; `source` is the
                reference original (encoder v1 is auto-selected).
            video_intelligence — `mode` one of description, categorization,
                moderation, search, custom (NOT `features`). Source must be
                https:// and meet duration minimums (description etc. ≥10s,
                search ≥4s) — check via metadata or tell user if too short.
        Example vmaf entry:
            {
                "output": "vmaf",
                "distorted": "https://example.com/encoded.mp4",
                "destination": {"url": "<destination_prefix>/vmaf.json"}
            }
        Example HLS entry (params on `stream[]`, not on the format object):
            {
                "output": "advanced_hls",
                "segment_duration": 6,
                "stream": [{
                    "video_codec": "libx264",
                    "audio_codec": "libfdk_aac",
                    "resolution": 720,
                    "framerate": "30",
                    "keyframe": "60",
                    "optimize_bitrate": 1,
                    "audio_bitrate": 128
                }]
            }
        Example video_intelligence entry:
            {
                "output": "video_intelligence",
                "mode": "description",
                "destination": {"url": "<destination_prefix>/vi"}
            }
    payload: optional opaque callback tag echoed by Qencode. Must be a
        string (or omitted). A JSON object is not submitted — the tool
        returns `{"error": "..."}` asking you to retry with a string.
        It is not the job body (`outputs` is). Do not put `source` or
        `format` in `payload`.

`encoder_version` is injected automatically when omitted: `2` by default,
`1` when any output is `vmaf`. Stitch jobs (multi-source `stitch` array)
are not supported here — use `start_encode2_raw` with `encoder_version: 2`
per `qencode://recipe/stitching`.

Other composition defaults in this server's instructions (libfdk_aac,
optimize_bitrate, per-stream ABR params, etc.) still belong in each
`outputs[]` entry — consult the matching recipe via
`search_qencode_docs` + `fetch_qencode_doc` before submitting.

After this returns a `task_token`, in the SAME reply call `list_jobs`
with that token. If the job is already Done with a playable video URL,
also `open_player`. If the deliverable is a json/txt/srt/vtt file, call
`fetch_job_result` and give the user both the URL and the extracted
content.