Fixes #1698.

Problem

The SQL formatter mangled a UNION (or INTERSECT/EXCEPT) inside a derived table or CTE: it collapsed the nested indentation, put blank lines around the set-operation keyword, and appended the closing ) directly to the last inner SELECT (...current) AS final).

Root cause

indent was a bare counter, and handleSetOperation assumed set operations are always top-level: it did indent = 0 and clauseStack.removeAll() on every UNION/INTERSECT/EXCEPT. Inside a subquery that destroyed the .subquery frame, so the next SELECT formatted at indent 0 and the closing ) fell through handleCloseParen's block branch into the bare-append fallback.

Fix

Refactor to a nesting-derived model rather than patch the one symptom:

• indent is now a computed property equal to the number of open block frames (subquery / CREATE TABLE body). Every manual indent += 1 / -= 1 / = 0 is gone, so indentation can't be clobbered by clause handling.
• handleSetOperation pops only the current SELECT's clause frames back to the enclosing block, keeps the block frame, emits the keyword at the real nesting indent, and uses blank-line separation only at the top level (single newlines when nested).
• Per-statement state (including selectColumnIndentStack) is cleared on ;, fixing a stack imbalance.
• Oracle MINUS routes through the same path.

For the reported query the formatter now indents the derived table one level, aligns both UNION blocks, and puts ) AS final on its own line.

Tests

13 new cases in SQLFormatterServiceTests: UNION / UNION ALL / INTERSECT / EXCEPT in a derived table, UNION in a CTE, UNION in a nested subquery, chained top-level unions, top-level INTERSECT, INSERT ... SELECT, window-frame ROWS BETWEEN stays inline, and two idempotency checks. The full suite (49 tests, including every pre-existing case: union, subqueryInFrom, cte, createTable, windowFunction) passes with no regressions.

Notes

• Output uses TablePro's existing house style (2-space indent + column river), matching the existing subqueryInFrom/cte tests, not the 4-space style in the issue's "expected" block.
• swiftlint --strict clean on the formatter source.
• Out of scope (would need the parser/AST rewrite): reflowing scalar subqueries inside the SELECT column river, ON CONFLICT / ON DUPLICATE KEY UPDATE bodies, and MSSQL [bracket] identifiers (conflicts with Postgres array subscripts).