Support DuckDB syntax: FROM-first, SEMI/ANTI joins, COPY, ATTACH, MACRO, PRAGMA and DuckDB 2.0 additions - #2643

DuckDB allows SEMI/ANTI JOIN without a LEFT prefix, and ANTI as a join
type at all. ANTI previously parsed silently as a table alias, producing
a wrong AST from valid SQL. Adds Join.isAnti(), the LEFT/RIGHT/FULL
SEMI|ANTI prefixes and the standalone form, keeping ANTI usable as an
identifier outside a join prefix.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

DuckDB parameterises a type with other types, e.g. MAP(VARCHAR, INTEGER).
Type arguments previously accepted identifiers and literals but not data
type tokens.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

DuckDB matches an INSERT's source columns either positionally or by
name. BY previously bound as a table alias, swallowing the clause.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

DuckDB writes the sample size first and the method behind it:
USING SAMPLE 10%, USING SAMPLE 10 PERCENT (bernoulli),
USING SAMPLE 10 ROWS (system, 377), USING SAMPLE reservoir (50 ROWS).
The method is now optional, RESERVOIR is a known method, and a bare
sample size accepts a unit.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

DESCRIBE now accepts a query as well as a table, so DESCRIBE SELECT ...,
DESCRIBE FROM ... and DESCRIBE WITH ... parse. The described query is
reachable from TablesNamesFinder and rendered by the deparser.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds PRAGMA name, PRAGMA name(arguments) and PRAGMA name = value, as
used by DuckDB and SQLite to read and set database settings. PRAGMA
stays a non-reserved keyword.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds [FORCE] INSTALL extension [FROM repository] and LOAD extension,
where the extension and repository may be an identifier or a string
literal. INSTALL and LOAD stay non-reserved keywords.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds CONNECT target [AS alias] and DISCONNECT [name], which route a
session's queries to a remote database. Oracle's CONNECT BY is
unaffected.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds PREPARE name AS statement and DEALLOCATE [PREPARE] name. EXECUTE of
a prepared statement already parsed. The prepared statement is a nested
Statement, so TablesNamesFinder reaches its tables.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds COPY table [(columns)] FROM path [(options)] and
COPY {table | (query)} TO path [(options)]. Options such as
FORMAT PARQUET or AUTO_DETECT true are free-form keywords, so they are
collected verbatim. COPY stays valid as MySQL's ALGORITHM value.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds CREATE [OR REPLACE] [TEMPORARY] MACRO name (parameters) AS body for
both scalar macros and table macros (AS TABLE SELECT ...). Parameters may
carry a default value, as in add(a, b := 5).

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds CREATE EXTENSION REPOSITORY [IF NOT EXISTS] name [WITH PREFIX
prefix], which registers a custom repository to install extensions from.
PostgreSQL's CREATE EXTENSION is unaffected.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

DuckDB writes CREATE TRIGGER ... REFERENCING OLD TABLE AS o NEW TABLE AS
n FOR EACH STATEMENT <statement>, i.e. the standard header with a SQL
statement instead of EXECUTE FUNCTION. The standard-form parser now
accepts a statement body, and such triggers are recognised and rendered
as standard rather than MySQL-style triggers.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds the simplified statement form UNPIVOT tbl ON columns
[INTO NAME name VALUE values] as UnPivotQuery, the counterpart of the
existing PivotQuery. The FROM-clause form UNPIVOT (...) is unaffected.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds JOIN t APPROX NEAREST n BY SIMILARITY expression, the vector
similarity join of DuckDB 2.0. APPROX no longer binds as a table alias
in that position. Also adds the lookahead the UNPIVOT statement's INTO
clause needed to keep the grammar free of choice conflicts.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2

Adds WITH RECURSIVE tbl (a, b) USING KEY (a, avg(b)) AS (...), which
aggregates a recursive CTE by key instead of accumulating every row.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EqinwHBKuAtPEmtXr3b5P2