Map: core / syntax

core/src/syntax/ turns source text into the surface AST — the only place that sees raw bytes and bare words.

• lexer.rs — lex(source) -> Result<Vec<(Token, Span)>, LexError>; the Token enum. An unterminated string names the exact closer it still wants, StringForm::closing() — a bumped literal wants its # run back after the ', so the bare reflex never closes it. Folding fd 1 onto fd 2 (1>&2, and >&2, which is the same redirect spelled short) is refused at the lexer, pointing at the warn builtin, and at 2>&1 for a program holding the direction backwards: a diagnostic is a verb here, not a second name for the byte channel.
• parser.rs — parse(source) -> Result<Vec<Stmt>, ParseError>; parse_with carries a FileId. A trailing & is a stage terminator only so its refusal can name spawn { … }, whose handle you await. A digit glued to a comparison operator inside $[…] ($[2>3], lexed as the file-descriptor redirect 2>) earns a diagnostic that names the shape and asks for spaces, not a bare “redirect” token. << is the here-string redirect; the bash spellings (<<<, a glued heredoc <<EOF) earn targeted diagnostics naming ral’s form. A run of separators collapses to one token, Token::Semi when it contains a ; and the soft Token::Newline otherwise: a ; is hard, never crossed by a pipeline or chain continuation. parse_pattern is the one place a duplicate binder is refused — a pattern binds all its names or none, so let [same, same] = [1, 2] is a parse error at the pattern’s span, while a repeat across curried parameters stays ordinary shadowing.
• ast.rs — the surface AST. Ast is the expression node and Stmt the statement node; the enum is deliberately wide and flat (ast-stays-flat). Ast::Unit is the () literal — punctuation denoting the unit value, like [] and [:], so not a word. Ast::Case carries arms: Vec<CaseArm>, a finite list of tag-and-body alternatives the parser hands on whole (case-is-syntax-try-is-not). WordLiteral::classify is the numeral grammar: purely lexical, never a round trip through printing, so 007 and 1.50 are numerals while 1e5, which merely happens to f64-parse, is text (a-numeral-denotes-its-number).
• group.rs — pre-pass detecting mutually recursive binding groups, consumed by the elaborator.
• quote.rs — bare-word classification (is_bare_word, quote_word, quote_word_if_needed), shared with the REPL. It is the dual of WordLiteral::classify: a string may go bare only where the numeral grammar declines it, since a bare 007 would read back as the number 7.
• tag.rs, free_refs.rs — variant-label tagging and free-reference scans.

Recursion is bounded by a single shared depth cap, so adversarial nesting rejects cleanly instead of overflowing the stack. One NESTING_DEPTH_LIMIT governs both stages.

• The lexer enforces it at scan_token_group, the sole recursion through bracketed token groups.
• The parser’s three mutually-recursive sub-grammars each descend through exactly one guarded chokepoint, all counting against Parser::nested’s shared depth: values through parse_primary, arithmetic through parse_expr_atom (the unary prefixes - / not and parenthesised sub-expressions bottom out here), patterns through parse_pattern (list and map patterns recurse back through it per element).

The five reserved control operators are recognised at this layer; everything else is a library binding. See docs/SPEC.md §3, §17.1 for the grammar.