[kvlonge]

Follow-up to #89 (D186). It is based on #89's branch fix/d150-v3, so the diff shows only these two commits. Once #89 lands, I'll rebase this PR onto main, retarget it to main, re-derive the corpus rates and gate it.

Commits

1. A bare Snowflake LIMIT or OFFSET is an alias, as Snowflake reads it (import 163, wip/syntax-evidence 462b572).
   • SELECT 1 LIMIT names its column LIMIT.
   • In FROM t LIMIT, LIMIT names the table, so LIMIT.a reads its column.
   • A bare LIMIT is an alias when ;, ), WHERE, ORDER BY, another LIMIT or the end of input follows it.
   • A bare OFFSET is an alias when ;, ) or the end of input follows it.
   • LIMIT OFFSET 1 stays a syntax error (1003). LIMIT with a row count, and LIMIT NULL, keep their row-limit meaning.
2. SQ0105's help for a long set-operation chain.
   • A flat UNION ALL chain of 513 or more branches is still refused with SQ0105. The depth limit stays because convert recurses once per branch and overflows its stack at 513 branches.
   • When the statement holds more than 250 set operators, the help now says to put at most 250 branches in each common table expression and combine them with the same operator. Other deep statements keep the general help.
   • An iterative convert for long chains is on the product lane's list.

Evidence

• The parse verdict changes for none of the 56,748 receipted statements.
• The everyday corpus query SELECT order_id FROM orders LIMIT moves from refused SQ0003 to translated, with the alias quoted as "LIMIT". The translation rates move to 288 of 436 (strict) and 357 of 436 (repository).
• 1,000 UNION ALL branches, split into four CTEs of 250 as the new help says, pass both check and convert.
• The test suites, clippy, fensu (0 faults) and the hygiene scan pass.

[kvlonge]

Closing: SQLQuack lands through local-gate trains, not PR merges (D94), so open PRs only add clutter (D628/D630). The branch is kept. Any work here that isn't on main yet is tracked by the train owner and lands through a train.