While loops needed optional parens and struct-literal suppression: gate 154 -> 111 events #2588

#2587 found that 138 while loops never reached the output and that unused local variable was their symptom, not a corpus defect. This is the cause and the fix.

Two defects, one construct

Mandatory parentheses. parse_while_stmt began with self.expect(TokenKind::LParen)?. 115 of 586 while statements are written while i < 100 { without them, so the expect failed at the first token and recovery discarded the entire loop.

Struct-literal ambiguity. Making the parens optional was not enough: parse_expr then read while i < count { as count { ... } — a struct literal — and swallowed the loop body, leaving the following return at apparent top level. Rust has the same ambiguity and solves it the same way, with a flag that suppresses struct literals inside an unparenthesised condition.

Zig requires the parentheses, so the emitter writes them regardless of how the spec spelled it: while (i < 100) {.

Measured

spec parse gate:      154 -> 111 recovery events, 0 blind   (ratchet holds)
while loops emitted:  286 -> 302
specs losing a loop:   51 -> 38
zig ast-check valid:  204 -> 202
total errors:         669 -> 676

The gate improvement is the headline: 43 fewer recovery events, the largest single move it has made in 48 iterations of holding at 154.

The two lost specs were passing on deleted code

stdlib_tb.t27 and vcd_trace.t27 were VALID with zero errors. Both had their loops dropped, and the dropped loops took the problematic code with them. Now the loops are emitted and their contents are checked: undeclared identifier 'result' and local variable is never mutated x4.

Same argument as #2584: a file that passes because its broken part was deleted is not passing. Two specs now fail honestly rather than passing on absence.

The intermediate state, recorded because it was tempting

With parens optional but the ambiguity unresolved: 218 valid, 615 errors — far better on both counters than what landed. And the gate FAILED, because 42 specs recovered more than their baseline while the total improved.

That state was the best-looking measurement of the night and it was wrong: the loop bodies were being consumed as struct literals, so the parser was skipping more code, not less. Two counters said yes and the gate said no, and the gate was right.

Boundary

• specs/fpga/testbench/stdlib_tb.t27
• specs/fpga/vcd_trace.t27