Add lean exporter

[max-rosenblattl]

♻️ Current situation & Problem

There is no way to persist a CaptionResult to disk.

⚙️ Release Notes

• Adds exporters/ package with lean.py: write_caption_result(result, path, *, rows_per_shard=4096, compression="zstd")
• Sharded Arrow IPC layout (recordings_{i:04d}.arrow), one shard per rows_per_shard recordings — ~32 shards for the full MHC dataset at the default, a single shard for typical small exports
• Embeds sensortslm.schema_version in each shard's schema metadata — no separate manifest file
• Arrow IPC over Parquet: matches the existing MHC input format, enables zero-copy mmap reads downstream, and our access pattern is whole-row (parquet's column-pruning advantage doesn't apply)
• Read path intentionally out of scope — a dedicated dataset class will live with the training pipeline; this module is write-only by design

📚 Documentation

Public API is documented inline. Usage:

from pathlib import Path
from exporters.lean import write_caption_result

write_caption_result(result, Path("exports/my_run"))
# → exports/my_run/recordings_0000.arrow, recordings_0001.arrow, ...

The exporters/ package uses an empty __init__.py so future format modules (e.g. exporters/timenet.py) can be added without forcing every consumer to load all formats' dependencies.

✅ Testing

Smoke-tested on the local MHC subset (../mhc_subset_testdata_hf) in the sensortslm 3.12 venv:

• Captioned 5 rows with statistical + structural + semantic extractors → 69 annotations
• Wrote to exports/lean_smoke with rows_per_shard=2 to force the multi-shard path → got 3 shards (~75 KB total)
• Verified all 13 expected columns including nested annotations struct list
• Verified sensortslm.schema_version=1 round-trips via feather.read_table(...).schema.metadata
• Verified Recording reconstruction (round-trip via a temporary read helper) preserves row_id, channel_names, values (NaN-aware), has_any_data, and all annotation fields

Code of Conduct & Contributing Guidelines

By creating and submitting this pull request, you agree to follow our Code of Conduct and Contributing Guidelines:

• I agree to follow the Code of Conduct and Contributing Guidelines.

[max-rosenblattl]

@coderabbitai review