fix: fetch(attrs, as_dict=True) excludes unrequested primary key

[dimitri-yatsenko]

Summary

Restores 0.14.x behavior where fetch('col1', 'col2', as_dict=True) returned dicts containing only the requested attributes.

In 2.x, proj() always includes primary key attributes, so the returned dicts contained extra keys the caller didn't ask for. This breaks downstream code that concats the result with an existing DataFrame that already has the primary key column — the concat creates duplicate columns and subsequent groupby fails.

The fix filters each dict after to_dicts() to include only the requested attributes. This is safe regardless of row order because we are removing keys from individual dicts, not rearranging rows.

Reproducing the issue

# Table with primary key (dataset) and secondary attrs (heading_dir, head_angle)

# 0.14.x — only requested attrs:
Table.fetch('heading_dir', 'head_angle', as_dict=True)
# [{'heading_dir': ..., 'head_angle': ...}, ...]

# 2.1.0 — primary key leaks in:
Table.fetch('heading_dir', 'head_angle', as_dict=True)
# [{'dataset': ..., 'heading_dir': ..., 'head_angle': ...}, ...]
# pd.concat with a df that already has 'dataset' creates duplicates

Test plan

• fetch('attr', as_dict=True) returns dicts without primary key
• fetch('KEY', 'attr', as_dict=True) still includes primary key (explicitly requested)
• Integration tests pass

[dimitri-yatsenko]

Won't fix — the 2.x behavior of always including primary key in fetch(attrs, as_dict=True) is intentional. It preserves row identity. Downstream code should handle column deduplication when concatenating.