An MLT property column has one type for the whole layer, while MVT values
are individually typed. Converting each value on its own left the encoder
to reconcile the mismatch, and its fallback for a column holding both
integers and doubles is to encode the whole column as strings, so an
attribute like gdp_md_est in tests/ne_110m_admin_0_countries turned into
values like "904.200000".
Summarize each attribute across the layer first and pick one type that can
hold all of its values, so mixed integer and floating point columns become
doubles. Only columns that mix numbers with strings, or booleans with
numbers, still fall back to strings, which is as close as MLT's typed
columns can get.
Also stop encoding JSON-object-valued attributes as MLT struct columns.
Struct children can only be strings, and a struct column is flattened into
"column name + child name" when it is read back, so an attribute `meta`
holding {"en": "one", "de": "eins"} decoded as separate `metaen` and
`metade` attributes. Nested JSON now stays JSON text, the way MVT
carries it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LxhELiLtpUFwrPSWwTYdHG
Claude