cmd_inputs: detect and regenerate stale .inputs metadata

When a gate's .weight tensor was rewritten by a later build pass (most
notably the bit-cascade comparator and modular ternarization work),
its existing .inputs entry from an earlier seed file no longer matches
the new fan-in. build_inputs now compares each existing .inputs length
against its corresponding .weight (for single-gate tensors with
weight.dim() == 1) and regenerates when they disagree, rather than
silently keeping stale routing.

Packed multi-gate tensors (weight.dim() > 1, e.g. memory.read.and)
use a different routing convention and are left alone.

This fixes the inconsistency that downstream tools like
safetensors2verilog were tripping on; on neural_alu8.safetensors,
15,264 stale entries get regenerated. The remaining ~3.4k gates
whose new naming patterns infer_inputs_for_gate doesn't yet recognize
end up with no .inputs at all (rather than a wrong-length one), which
is the cleaner failure mode.

build.py

@@ -2977,12 +2977,24 @@ def infer_inputs_for_gate(gate: str, reg: SignalRegistry, tensors: Dict[str, tor
 def build_inputs(tensors: Dict[str, torch.Tensor]) -> tuple[Dict[str, torch.Tensor], SignalRegistry, dict]:
     reg = SignalRegistry()
     gates = get_all_gates(tensors)
-    stats = {"added": 0, "skipped": 0, "empty": 0}
+    stats = {"added": 0, "skipped": 0, "empty": 0, "regenerated": 0}
     for gate in sorted(gates):
         inputs_key = f"{gate}.inputs"
+        weight_key = f"{gate}.weight"
         if inputs_key in tensors:
-            stats["skipped"] += 1
-            continue
+            # Detect stale .inputs (length doesn't match the gate's fan-in)
+            # for single-gate tensors and regenerate them. Packed multi-gate
+            # tensors have weight.dim() > 1 and use a different convention,
+            # so we leave their .inputs alone.
+            existing = tensors[inputs_key]
+            weight = tensors.get(weight_key)
+            if (weight is not None and weight.dim() == 1
+                    and existing.numel() != weight.numel()):
+                del tensors[inputs_key]
+                stats["regenerated"] += 1
+            else:
+                stats["skipped"] += 1
+                continue
         inputs = infer_inputs_for_gate(gate, reg, tensors)
         if inputs:
             tensors[inputs_key] = torch.tensor(inputs, dtype=torch.int64)