Real-corpus friction (Egocentric-10K/100K): no WebDataset ingest path; the video adapter's JPEG detour costs 2.3x landing size and a double lossy transcode for H.265 sources

## Context

I ran HFlow against the shape of `builddotai/Egocentric-10K` and `builddotai/Egocentric-100K` (the largest real egocentric factory corpora on Hugging Face, and the same source the `examples/egocentric` quickstart shard comes from). Both corpora are WebDataset-style tars of ~180 s H.265/MP4 clips (1080p30 for 10K, 456x256 for 100K) plus per-clip JSON sidecars and per-worker `intrinsics.json`.

The finding: **there is no first-party ingestion path for this corpus shape, and the only available video-to-MCAP adapter detours through JPEG, which costs roughly 2.3x landing size and a double lossy transcode for H.265 sources.**

## What exists today

- `src/hflow/importers/` contains only the LeRobot v3 converter. WebDataset-style tars have no importer; users must write their own converter (the documented escape hatch) or reach for `hflow.testing.write_video_episode`.
- `write_video_episode` is explicitly a demo/test helper, and it decodes the source video and re-encodes **every frame as JPEG** `sensor_msgs/CompressedImage` (`_build_video_camera_messages` in `src/hflow/testing.py`). The canonical transform then re-encodes that JPEG into in-band H.264.

So for an H.265 corpus the media is transcoded **HEVC -> JPEG -> H.264** (two lossy steps, one avoidable) even though H.264 is the canonical target and ffmpeg decodes HEVC directly.

## Measurements

Profile-matched HEVC media (ffmpeg `testsrc2`, same resolution/fps/codec as the corpora; the real footage is gated so I could not download it for this test). Apple Silicon macOS, ffmpeg 8.1.1 from Homebrew, hflow at `78a6b1f`.

30 s @ 1920x1080 @ 30 fps (Egocentric-10K profile):

| Step | Result |
| --- | --- |
| Source HEVC bytes (same span) | ~7 MB |
| `write_video_episode` landing MCAP (JPEG) | **55.3 MB** (~8x source bitrate) |
| Canonical episode H.264 media bytes | 24.5 MB (landing is 2.3x this) |
| Adapter wall time | 1.8 s |
| Cold full pipeline (`app.test`) | **12.0 s (~2.5x slower than realtime)** |
| of which `camera_frame_stats` decode | 8.1 s |

60 s @ 456x256 @ 30 fps (Egocentric-100K profile): cold full pipeline 2.9 s (~48 ms per source-second) -- the small profile is fine; the cost is concentrated at 1080p.

At 2.5x-slower-than-realtime, a single process QC pass over Egocentric-10K's 10,000 hours is on the order of a machine-year; the per-episode unit cost is what matters, and most of it is avoidable decode/transcode overhead plus JPEG round-tripping.

## Why it matters

1. **Disk**: wrapping a real corpus through the JPEG detour multiplies landing storage by ~2.3x versus what the canonical stage needs, before any pipeline run.
2. **Quality**: every real frame passes through two lossy re-encodes; `camera_frame_stats` and any downstream consumer measure artifacts of the JPEG step, not just of the source.
3. **Awkwardness**: the JSON sidecars (factory/worker/duration/fps/codec) map naturally onto `episode/v1` metadata records and `intrinsics.json` onto calibration attachments, but there is no example or helper wiring a WebDataset tar into that, so every team rediscovers the adapter path on their own.

## Suggestions (non-prescriptive)

- A first-party WebDataset-to-canonical importer (or a documented example next to `examples/egocentric`) that transcodes HEVC -> in-band H.264 **directly**, skipping the JPEG landing detour.
- Or, if the JPEG path is intentional for test isolation, a note in the docs saying so and pointing real-corpus users at the converter how-to, so the demo helper does not become the de facto ingestion path.

Happy to split any of this into smaller scoped issues if maintainers prefer.

Start by reading src/hflow/importers/, src/hflow/testing.py, and the _build_video_camera_messages path, then inspect examples/egocentric and the documented converter escape hatch. The issue presents both a first-party WebDataset importer and a documentation-only alternative, so the intended scope needs maintainer agreement; done should be a decided ingestion or documentation path with coverage for the described corpus shape and its media handling.

Written by the indexing model from the issue text.