[philrz]

Fixes #7468

When the input to a SQL query has a named record type and another record type, the table's columns came from only one of the record types. #7548 fixed the error values in the original repro, but SELECT * still dropped a column, and a reference to the dropped column was a compile error.

With this PR, the table has the columns of both record types:

$ super -f bsup -o t1.bsup -c "type foo={a:int64} values cast({a:1},foo), {b:2}"
$ super -c "SELECT * FROM 't1.bsup'"
{a:1,b:null}
{a:null,b:2}
$ super -c "SELECT b FROM 't1.bsup'"
{b:null}
{b:2}

(Previously {a:1} / {a:null}, and "b" is not a column or table.)

The fuser keeps named types apart, so the fused type of this input is fusion(foo|{b:int64}), a union with two record types, and recordOf() in compiler/semantic/schema.go used only the first. This PR:

• fuses all of the union's record types in recordOf(), so the columns are the same as for the input without the named type
• updates the test from fix bug with projected missing under nullish #7548, which expected the dropped column, and adds cases for SELECT b and for two named record types (where tip kept only b)

The fix covers every way SQL gets its input: a file, a pipe into SELECT, a subquery, or a CTE. With -dynamic, column references like SELECT b already worked. Because such a table now has all its columns, a join can newly report an unqualified column as ambiguous, as it already did for the same input without the named type.

The patch was drafted with Claude Code. I reviewed and tested it, and ran it through Claude Code's /code-review high.

[mccanne]

This stuff is a bit fragile, but step in the right direction.