Boxplot produces invalid Vega-Lite: `type` encoding channel leaks into output

[cpsievert]

Boxplot queries produce Vega-Lite JSON that fails validation against the Vega-Lite v6 schema. This issue was discovered in posit-dev/querychat#232, but the issue is still present on main.

BTW, in a future version of the Python package, I plan on not validating JSON against the schema by default, so that alone would fix this problem, but I also figured it's still probably worth reporting.

Reproduce with the CLI

ggsql exec \
  "SELECT grp, value FROM (VALUES ('A', 1), ('A', 2), ('A', 3), ('B', 4), ('B', 5), ('B', 6)) AS t(grp, value) VISUALISE grp AS x, value AS y DRAW boxplot" \
  --reader duckdb://memory

Paste the output into the Vega-Lite Editor and note the warnings about validation

What's wrong

Two issues in the generated spec:

1. Each sub-layer's encoding contains a "type" channel, which is not a valid Vega-Lite encoding property:

   "encoding": {
     "type": {
       "field": "__ggsql_aes_type__",
       "type": "nominal",
       ...
     }
   }

   The boxplot stat's internal "type" aesthetic (used to tag rows as box/median/whisker/outlier) is correctly dropped from the data values but still emitted as an encoding.

2. The y2 encoding on whisker/box layers carries axis, scale, type, and stack — properties that y2 (a secondary position channel) does not accept.