notebooks: MCAP robotics DataFrame demo

[everettVT]

Summary

A walkthrough notebook for daft.read_mcap against a real robotics recording — a 1.19 GB MCAP from the DapengFeng/MCAP dataset on HuggingFace (FAST-LIVO/hku1: Livox LiDAR + IMU + stereo compressed cameras, 29,702 messages over 127.7s).

Covers the things a robotics user would actually ask of an MCAP DataFrame:

• Schema inference (topic / log_time / publish_time / sequence / data)
• Per-topic groupby for "what's in this recording"
• Topic filter pushdown — topics=[\"/livox/imu\"] avoids materializing camera payloads
• Time-window pushdown via start_time / end_time
• topic_start_time_resolver — the new per-file per-topic keyframe alignment from Daft #5886 (v0.7.2)
• A 50ms-bucket join between LiDAR sweeps and IMU samples — the time-aligned query that's painful in raw MCAP

Known limitation

Direct HTTP reads (daft.read_mcap(\"https://...\")) currently fail with TypeError: Expected a FileSystemHandler instance, got HTTPFileSystem — daft/filesystem.py wraps fsspec in PyFileSystem without going through FSSpecHandler. Tracked in Eventual-Inc/Daft#6983. Notebook uses `huggingface_hub.hf_hub_download` as the workaround; will update once #6983 lands.

Placement

Dropped in `notebooks/` alongside the existing format-walkthrough notebooks (`getting_started_with_common_crawl.ipynb`, `window_functions.ipynb`). Happy to convert to a PEP 723 script under `examples/io/` in a follow-up if that fits the repo direction better.

Test plan

• `jupyter nbconvert --execute` runs against a fresh `~/.cache/huggingface` (note: 1.2 GB download)
• Schema output matches the README claims in the markdown cells
• Topic counts: `/livox/imu` ≈ 25,871, other three ≈ 1,277 each
• Verify `daft[mcap]>=0.7.13` extras spec resolves (the install cell pins it)

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 30af30239a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Match adjacent buckets for ±50ms window

The bucket equality join only matches IMU/LiDAR rows that fall in the exact same 50ms epoch bucket, which misses valid pairs near bucket boundaries even when they are within ±50ms (for example, timestamps 2ms apart on opposite sides of a boundary). Because this section claims to return IMU samples within ±50ms of each LiDAR frame, the current logic silently drops qualifying matches and can skew downstream alignment/fusion analyses.

Useful? React with 👍 / 👎.