fix(checks): skip empty state streams in timestamp sync offsets (#691)

[shobhitagnihotri69]

Summary

Fixes #691. Mirrors the empty camera guard introduced in #668 by ensuring candidate reference state streams actually have recorded timestamps (stamps_ns.size > 0) before picking a sync reference and computing start/end offsets.

Problem

When an MCAP file has summary metadata indicating messages for a non-camera topic (message_count >= 2), but zero actual records are written to disk, timestamp_regularity picked that empty state stream as the reference stream and crashed with IndexError: index -1 is out of bounds for axis 0 with size 0 on ref_stamps[-1].

Changes

• In _timestamp_regularity_keys: Filter state_topics to only those where episode.channel(topic).timestamps.size > 0.
• In _measure_timestamp_regularity: Filter state_topics to only those where per_topic[topic].stamps_ns.size > 0.
• If no populated state stream exists, cross-stream sync offsets are skipped gracefully. If other populated state streams exist, max() selects the densest populated state stream instead of the empty one.
• Added comprehensive unit tests in tests/test_checks.py.

All 2,478 tests pass and linting checks are clean.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Filters empty state streams from timestamp sync reference selection.

The PR appears safe to merge.

What we checked:

• Predicted sync keys differ from results: Both paths select the same topics and test the same channel timestamps. The measurement then writes the keys returned by the prediction function.

Summary

This PR updates timestamp_regularity to choose sync references only from state streams with recorded timestamps. If no selected state stream has timestamps, it still reports per-topic sample counts but emits no sync offsets.

• Camera sync offsets now use only state streams with timestamps.

_{Reviews (1) · Last reviewed commit: "fix(checks): skip empty state streams in..."}

[kstonekuan]

LGTM, merging.