v2 parse: a '-' that begins a line after an operand refuses (NewlineBeforeDualRoleOperator), as the seed refuses it (PR2a, stacked on #12773)

[gunbai-bot]

Stacked on #12773 (base: its branch). Discharges the trigger of the #12769 row, and is the production consumer of #12773's RefuseOnMatch / AfterLineBreak arms and its stream layout.

The defect, now on main

- is both infix (subtraction) and prefix (negation), and the dag grammar has no statement terminator. So x followed by a line - 1 is either one subtraction or two statements. The seed refuses it as ambiguous (v1.compiler.parse is_ambiguous_prefix_infix_newline_boundary, whose dual-role set is exactly -). v2 read it silently as one subtraction. Since #12817 made block forms operands of binary_expr, a block statement followed by a - line merges the same way: silent wrongness, below the floor.

The change

v2.extdeps.languages.dag dag_grammar_additive_expr_helper: at every additive continuation, the first alternative is
RefuseOnMatch { AfterLineBreak { - }, NewlineBeforeDualRoleOperator }.

• A - preceded by a line break refuses the parse at the -, with a typed reason from the closed v2.std.parse_refusal_reason vocabulary (new variant NewlineBeforeDualRoleOperator).
• A same-line - falls through to subtraction.
• Every expression reaches this continuation, so the rule holds for every statement form, matching the seed's rule.

Controls (v2.test.parse.newline_dual_role_operator_parse), green on a remote run

• x ⏎ - 1 refuses with NewlineBeforeDualRoleOperator at the - (bytes 51..52).
• match w {..} ⏎ - n refuses at the - (bytes 101..102).
• x - 1 on one line parses and normalizes.
• Emit side (DESIGN §4, one grammar both ways): the refusal is ingest-only (RefuseOnMatch derives no tree, and emit reads formal productions). What emit must guarantee is that it never writes what ingest refuses. Emitted text breaks a line only where a token class's emit transform is EmitLineBreaks, and no dag token class has one, so the dag emitter can never write a line-break-preceded -. no_dag_token_class_emits_a_line_break_holds reds the moment a dag class gains a line-breaking transform.

Mutation (test.claim.long.newline_dual_role_operator_mutation_witness_test), green

• The witness rebuilds the live dag grammar with every RefuseOnMatch arm replaced by a never-matching terminal.
• It asserts that such arms exist and that every one is the newline refusal. There are several because the additive row is a helper inlined at each operand position of the comparison and equality rows, but it is one authored row.
• Under the mutant both next-line forms parse again (the silent merge returns); under the live grammar both refuse; the same-line form parses under both.

Census

Over the 190 corpus files with a block-headed operand site, run with v2.compiler.parse_acceptance_census: 0 refuse with NewlineBeforeDualRoleOperator. A static scan finds no block-headed statement followed by a - line in dag/ or src/v2. So nothing in the tree changes meaning; the construct becomes unwritable from here on.

Rung

The #12769 class goes from silent wrongness to structurally guaranteed: no Accepted program contains a line-break-preceded dual-role - at an operand boundary.

🤖 Generated with Claude Code