dag grammar: node let-sugar guard is its whole head (fix forward #13056)

[gunbai-bot]

Fixes forward #13056, which merged with a parse regression reported by tidy-raven-393 and calm-boar-904.

Defect

dag_grammar_let_sugar_head_expr guarded the expression statement with !(binding_name "=" | "node" binding_name). The node let-sugar is node name (: | =) expr, so the guard was one token short. Statements of the following shapes start with node <name> without being sugar:

• a bare node statement followed on the next line by a statement that starts with a name;
• an arm => node followed by the next arm;
• node followed by if.

The guard excluded these from the expression alternative, and the sugar also refused them. The result was parse_g0_tokens_remain.

Fix

The guard is now the sugar's whole head: "node" binding_name (":" | "="). That is three tokens, inside the lookahead cap of 3, so the choice-overlap proof still reads it and the roster stays at 0.

Controls

v2.test.parse.bare_node_statement_route. Each claim is a route claim reading the stored Bool from the warm bns_verdicts:

• tidy-raven-393's fixtures a1 (arm => node before the next arm) and a3 (node before if), verbatim. Each must parse with zero let_expr nodes, i.e. it takes the expression route.
• node y = 1 must still take the let route (exactly one let_expr).

These are red-first: the two fixture claims fail on main and pass here. The sugar claim passes on both. (Red/green for these claims: from this PR's CI floor run.)

Corpus

A scan of dag and src/v2 for a statement node at end of line followed by a name-led line finds 6 files. Each is parsed whole with the dag grammar, on main and on this branch:

• src/v2/std/node.dag (N7 closure)
• dag/gunbc/scm/object_store.dag
• src/v2/workflow/realization_runner.dag
• src/v2/test/claim/wrap_decision_predicate_test.dag
• src/v2/test/claim/verilog_interlock_emission_test.dag
• src/v2/std/model_core.dag

Measured (one remote dispatch, same binary, each file's whole text parsed as a module with the dag grammar): all 6 refuse on main, all 6 parse on this branch.

Why #13056's evidence missed it

• The 32-file parse was a sample, not the corpus. It selected the files carrying the sites the change was meant to touch. None of the six files above was in it: a whole-file parse of any of them refuses at the guarded head. A whole-corpus parse would have caught this; the sample could not.

• The census enumerated sites whose meaning the change was meant to alter. It never enumerated what each new guard excludes. A guard must not exclude any input the sibling alternative fails to accept: if it does, it removes inputs rather than disambiguating them. Here the guard excluded node name while the sibling only accepts node name (: | =).

• What evidence would have caught it:

  • a guard-soundness check per not-predicate: the guard's strings are a subset of what the sibling alternative can start with and continue at the guard's depth;
  • and, as backstop, the whole-corpus parse differential (accepted on main → accepted on the head).

  I'll propose the first as a check in the prefix-table PR that follows, since it reads the same per-production prefix table.

🤖 Generated with Claude Code