A signed right shift needs three angle brackets

bootstrap/src/compiler.rs

@@ -16668,6 +16668,25 @@ impl VerilogCodegen {
                         // in $signed() (Verilog sign-extends it into the wider signed
                         // context) and the unsigned operand is zero-extended one bit,
                         // $signed({1'b0, x}), so it stays non-negative.
+                        // W-27: a RIGHT SHIFT of a signed value needs `>>>`.
+                        // Verilog's `>>` fills with zeros whatever the operand
+                        // is declared, and the repair below fires only on
+                        // ordered relations, so no shift was ever fixed up. On
+                        // `specs/igla/race/cordic_fixed.t27:42`, `x - (y >> shift)`
+                        // with x=100, y=-64, shift=2 simulates to -16268 where
+                        // the spec, C and Zig all say 116: y = 16'hFFC0 = 65472,
+                        // 65472 >> 2 = 16368, 100 - 16368 = -16268.
+                        //
+                        // This project's own hand-written golden CORDIC RTL
+                        // writes `y0 >>> 1` for the identical rotation, so the
+                        // right operator was known and the expression path
+                        // emitted it zero times across 559 generated modules.
+                        let shift_signed = node.extra_op == ">>"
+                            && matches!(
+                                self.expr_width_signed(&node.children[0]),
+                                Some((_, true))
+                            );
+                        let op = if shift_signed { ">>>" } else { op };
                         let ordered_rel = matches!(op, "<" | "<=" | ">" | ">=");
                         let rel_signed = ordered_rel
                             && match (

bootstrap/stage0/FROZEN_HASH

@@ -1 +1 @@
-d08617f241b7aebdbef21636a651180b9b303ea20c1077452be83165e3a5b736
+b109d745f01460f08fb2f8547d27a816dfecaff6c3f917da90f2f465c3daa86d

docs/now/2026-08-30-a-signed-right-shift-needs-three-angles.md

@@ -0,0 +1,12 @@
+# NOW -- A signed right shift needs three angle brackets (2026-08-30)
+
+## The Verilog backend filled with zeros where the spec said sign (Refs #2860)
+
+- `>>` was mapped to Verilog's `>>` unconditionally; Verilog's `>>` is LOGICAL and fills with zeros however the operand is declared
+- the signedness repair the backend already has fires only on ordered relations `< <= > >=`, so no shift was ever fixed up
+- simulated on the ACTUAL generated module: `cordic_x_next(100, -64, shift=2)` returned **-16268**; the spec, C and Zig all say **116**
+- y = -64 is 16'hFFC0 = 65472; 65472 >> 2 = 16368; 100 - 16368 = -16268
+- this project's own hand-written golden CORDIC RTL writes `y0 >>> 1` for the identical rotation, so the right operator was known and the expression path emitted it ZERO times across 559 modules
+- `>>>` occurrences corpus-wide: 2 -> 369 across 48 specs; both of the old two were inside string literals
+- iverilog acceptance unchanged 373/559 -- the wrong shift was always legal Verilog, which is why no gate could have caught it
+- tempered, and the audit said so first: 42 of the 45 affected specs shift non-negative values today, so they are LATENT. The three igla/race CORDIC specs are value-corrupting now