[kyleconroy]

Moves the pin from the alpha nightly (8cbdaba6ac, v2.0.0-alpha38195) to the v2.0.0 tag (d54cbbe3df). It also rehearses the regeneration workflow (PLAN.md § Milestones, item 7).

Gates (all against the v2.0.0 oracle)

• TestCorpus: the regenerated corpus has 56,064 cases over 4,925 files (54,106 accept, 1,958 reject; 11 dropped for oracle timeouts). darkwing agrees on every case except post-parse rejects, which are recorded as todo/skip entries the same way as before.
• TestCorpusErrorMessages: every rejection matches the oracle's first line verbatim.
• serialize-diff -corpus: 31,515 match, 0 mismatch, 103 skipped. Goldens are regenerated (2,626 statements) and TestSerializeGoldens passes.
• Dev loop: go build ./... && go vet ./... && gofmt -l . && go test -race ./... is clean. DARKWING_DUCKDB_SRC=… go test ./internal/grammar checks darkwing's assembled grammar against upstream's inlined_grammar.gram.

Oracle

• Built from source. No v2.0.0 CLI binaries are published yet. I built the CLI from the tag with bundled_extensions.cmake and DUCKDB_VERSION=v2.0.0. It reports v2.0.0 (Cyanoptera) d54cbbe3df, the same commit as the PyPI duckdb==2.0.0 wheel.
• Library grammar, not the CLI's. The v2.0.0 CLI opens with active_grammar_extensions=['shell'], which re-adds bare expression statements (1 + 1). The library grammar rejects them, and so does the Python library. The oracle and serialize-diff now run SET active_grammar_extensions=[].
• Agent mode off. The new CLI switches to an agent output format when it detects a coding agent in the environment, so the oracle sets DUCKDB_AGENT_MODE=0.
• Release URL. duckdbsrc.NightlyURL is replaced by ReleaseURL, and the difftest lane downloads from it. The lane will fail at download until DuckDB publishes the release assets.

Port

• Grammar: re-vendored verbatim, including the new bare_label_keyword.list and external_resource_source.gram. .gitignore's duckdb pattern was also hiding new files under internal/grammar/duckdb/; it's now anchored to /duckdb.
• Engine:
  • Tokenizer: backtick identifiers, operator lexing (:, ->, **, //, trimming), stricter digit separators, unterminated block comments, string-literal continuation across lines, raw dollar-quoted text.
  • Grammar loader: ordinary PEG precedence (upstream's new recursive-descent rule parser). The old "/ binds one element" quirk is gone, and CLAUDE.md is updated to say so.
  • Matchers: identifier, number and operator matchers; bare-label and AnyOperatorLiteral overrides; the new packrat list.
• Statement transformers: the changes to CREATE/ALTER/DROP, DML, COPY and generic options, DESCRIBE/SHOW and the misc statements. Adds the new extension-repository and external-resource statements, with upstream's new Parser Errors copied verbatim.
• Query/expression transformers: the new expression precedence (IS/NOT/IN, Infix/CustomPrefix operators), literals, nested qualification paths, window and cast errors, FETCH/LIMIT forms, PIVOT, and an expression-depth limit that matches the oracle's 1000 boundary.

Known gaps (not in this PR)

• Unicode whitespace (predates this PR): darkwing's set differs from upstream's Parser::StripUnicodeSpaces. It rejects U+200B, U+2060 and U+FEFF, and accepts U+0085, U+1680 and U+2028/2029. The corpus doesn't exercise this. A faithful fix rewrites the query before parsing and shifts error offsets, so it needs a decision.
• Alias-error rendering: "Table reference … cannot have two aliases" shows source text where upstream shows its re-rendered SQL. They differ only for alias: (VALUES …) AS u.

---

Generated by Claude Code