Fix case-sensitive VISUALISE column reference validation

[cpsievert]

Summary

SQL treats unquoted identifiers as case-insensitive — SELECT REGION FROM t works even when the column is stored as region. However, ggsql's VISUALISE clause matches column references case-sensitively against the result schema. Since DuckDB lowercases unquoted identifiers in query results, writing VISUALISE REGION AS x fails when the result column is region:

aesthetic 'x' references non-existent column 'REGION'

Fix

Adds a normalize_column_names() step early in prepare_data_with_reader (before merge, validation, and query building). It uses case-insensitive matching to resolve VISUALISE column references to the actual column names in the result schema. This is reader-agnostic — it normalizes to whatever the reader returns, not specifically to lowercase.

Reproducing

CLI:

# Create a DuckDB file with lowercase-stored columns
duckdb repro.duckdb "CREATE TABLE t AS SELECT 'A' as category, 10 as value UNION ALL SELECT 'B', 20;"

# SQL resolves CATEGORY case-insensitively, but VISUALISE fails
ggsql exec --reader "duckdb://repro.duckdb" \
  "SELECT category, value FROM t VISUALISE CATEGORY AS x, VALUE AS y DRAW bar"
# => Validation error: Layer 1: aesthetic 'x' references non-existent column 'CATEGORY'

Python:

import ggsql

reader = ggsql.DuckDBReader("duckdb://memory")
reader.execute_sql(
    "CREATE TABLE t AS SELECT 'A' as category, 10 as value "
    "UNION ALL SELECT 'B', 20"
)

# SQL resolves CATEGORY case-insensitively, but VISUALISE fails
reader.execute(
    "SELECT category, value FROM t "
    "VISUALISE CATEGORY AS x, VALUE AS y DRAW bar"
)
# => Validation error: Layer 1: aesthetic 'x' references non-existent column 'CATEGORY'

Test plan

• Added test_case_insensitive_column_references test
• All 874 existing tests pass

[thomasp85]

This assertion is wrong. Each layer only takes from the global mapping what their data source and aesthetic requirements dictates. It's better to wait until global mapping has been merged into the layers and then do the normalisation only for the layers