feat(newline): 'terminator' mode, a NEWLINE at every line break

[theoephraim]

newline.mode: 'terminator': a NEWLINE at every line break

newline mode was built for dotenv-style grammars, but its NEWLINE is a statement separator: one token between two content-bearing lines, placed at the start of the later line, with blank and comment-only lines collapsed and nothing before the first or after the last line. A grammar whose AST keeps blank lines and comments as lines cannot be written on that. env-spec is one (File = many(Line), Line = Stmt NEWLINE? | NEWLINE), which is why it sits on indent mode today with many(alt(INDENT, DEDENT)) filler in every rule and an indent-stack tree-sitter scanner it never uses.

NewlineConfig.mode: 'terminator' emits one NEWLINE at every block-context line break, placed at the break (zero-width, before the \n / \r\n): blank lines, comment-only lines, leading and trailing breaks included; breaks inside flow delimiters stay suspended; a final line with no break has none. Positions are exact per line, so a blank-line node maps to its own line. It is also the shape the derived tree-sitter scanner already has (stateless, one NEWLINE per break where the grammar permits one), so parser and scanner agree more closely than in separator mode. 'separator' stays the default and is byte-identical.

Three emission points in gen-lexer.ts: a content line's break, a blank line, a whitespace-only (tab) line; the line-start boundary emission is skipped in terminator mode. The emitted engine inherits the mode through the createLexer fallback. The portable TS/Go/Rust newline lexers on portable-newline-mode will need the same three points once that branch lands; happy to follow up there.

Gate: test/newline-mode.ts sections 7 to 9. Section 7 pins the lexer (offsets for every case above, CRLF, flow suspension, no INDENT/DEDENT, separator mode unchanged) and the parser (a blank line is its own leaf at its own offset). Section 8 emits the grammar with jsTarget and checks the emitted engine parses identically to the interpreter, since the mode reaches it through the baked createLexer config. Section 9 runs the tree-sitter CLI when present and checks the derived parser produces one newline node per break for blank, leading, and trailing breaks, and none inside flow.

One thing section 9 turned up, outside this change: Line = Stmt NEWLINE? | NEWLINE (the natural shape for a grammar that keeps blank lines) is LR-ambiguous after a statement, so tree-sitter generate reports an unresolved conflict for it. The PEG engine resolves it greedily; tree-sitter needs a declared conflict or a prec on the optional. The section therefore checks the scanner claim on Program = many(NEWLINE | Stmt). Emitting that resolution from gen-treesitter (a trailing opt(X) whose X also starts a sibling alternative) looks like a small follow-up; happy to take it.

npm run check: 50/50.