fix: reject implicit ?T -> bool coercion instead of silent false (issue 0169)

The Optional->Concrete unwrap classify rule treated ?i64 -> bool as
unwrap+narrow (both builtin), silently yielding false for every optional
(present or null). specs.md defines no implicit optional->bool
conversion. Reject it: conversions.zig adds an optional_to_bool_reject
plan (dst == bool, child != bool); coerce.zig emits a located diagnostic
suggesting '!= null'. Covers arg/field-init/return via the shared
coerceMode. The if-opt presence test (issue 0164) is a separate path,
untouched.

Regression: examples/diagnostics/1199-diagnostics-optional-to-bool.sx +
conversions.test.zig unit test. Verified by 3 adversarial reviews, suite
789/0. Filed adjacent issue 0179 (whole implicit ?T->concrete unwrap
family silently miscompiles a null optional; design-touching).

src/ir/conversions.test.zig

@@ -48,6 +48,13 @@ test "conversions: classify covers the built-in coercion ladder" {
     try std.testing.expectEqual(Plan.optional_unwrap, cr.classify(opt_i64, .i64));
     try std.testing.expectEqual(Plan.void_to_optional, cr.classify(.void, opt_i64));
 
+    // `?T → bool` is NOT an unwrap-then-narrow presence test (issue 0169):
+    // it must reject, never silently produce `false`. But `?bool → bool`
+    // is a genuine unwrap of a bool payload.
+    try std.testing.expectEqual(Plan.optional_to_bool_reject, cr.classify(opt_i64, .bool));
+    const opt_bool = tt.optionalOf(.bool);
+    try std.testing.expectEqual(Plan.optional_unwrap, cr.classify(opt_bool, .bool));
+
     // Tuple → tuple, same arity.
     const t_ss = tt.intern(.{ .tuple = .{ .fields = &[_]TypeId{ .i64, .i64 }, .names = null } });
     const t_ii = tt.intern(.{ .tuple = .{ .fields = &[_]TypeId{ .i32, .i32 }, .names = null } });