feat(governance): add Gebru-style datasheets for training datasets

[Hopelynconsult]

Summary

• Adds governance/datasheet.py — Gebru et al. 2018, "Datasheets for Datasets". Companion to the Mitchell-style model cards from feat(governance): add automated model card generator #37.
• Adds scripts/generate_datasheet.py mirroring scripts/generate_model_card.py so the release CI pipeline can run them in sequence.
• Public API matches the model_card module (build_datasheet / render_markdown / write_datasheet / generate) so contributors only learn one pattern.

Why we need this

A model card describes the model. A datasheet describes the dataset that trained the model. The two artifacts answer different questions and the responsible-AI literature is clear that releases need both — for ClimateVision specifically, the dataset choices (which biomes, which years, which label sources) drive the geographic-bias risk that #30 and the bias-audit framework are trying to surface, so we need a structured place to document those choices alongside every release.

What's in the PR

Sections covered: motivation, composition, collection process, preprocessing, uses (intended + inappropriate), distribution, maintenance.

• Datasheet dataclass with JSON serialisation, mirrors ModelCard
• REQUIRED_QUESTIONS schema enforces minimum answers (purpose, creators, instances, labels, splits, source, timeframe, intended_uses, inappropriate_uses) — anything else is free-form so the schema can grow without code changes
• Sensible defaults for inappropriate_uses and maintenance so existing dataset manifests don't have to be exhaustive on day one
• 12 unit tests covering: build, override behaviour, all required-field validations, markdown rendering shape, JSON round-trip, and YAML manifest loading

Follow-ups (out of scope here)

• Author the first concrete datasheets for the Sentinel-2 deforestation, ice-melt, and flood training datasets (separate doc-PR per dataset).
• Wire scripts/generate_datasheet.py into the release CI alongside generate_model_card.py.
• Cross-link datasheet → model card in governance/model_card.py so each card includes a pointer to its training datasheet.

Test plan

• pytest tests/test_datasheet.py -q → 12 passed
• Reviewer: confirm REQUIRED_QUESTIONS is the right minimum bar — happy to relax/tighten before we author the first real datasheet.

🤖 Generated with Claude Code