placement: Add some extra checks to limit the risk of regressions.

Given noone understand this code, this will prevent some semi likely
regressions from going unnoticed. The main risk is the shims not
producing the exact same results as the real implementation and
cause different code paths to be taken.

As of this commit, both the "real" and "shim" implementation were given
the same set of tests with print() at every step of next_to. The
resulting log was then checksummed to ensure both are identical.

lib/awful/placement.lua

@@ -717,6 +717,10 @@ local function get_relative_regions(geo, mode, is_absolute)
         dgeo = geo.drawable.get_wibox():geometry()
     elseif geo.drawable and geo.drawable.drawable then
         bw, dgeo = 0, geo.drawable.drawable:geometry()
+    else
+        -- The placement isn't done on an object at all, having no border is
+        -- normal.
+        assert(mode == "geometry")
     end
 
     -- Add the infamous border size
@@ -1392,6 +1396,7 @@ function placement.next_to(d, args)
     args = add_context(args, "next_to")
     d    = d or capi.client.focus
 
+    local osize = type(d.geometry) == "function"  and d:geometry() or nil
     local original_pos, original_anchors = args.preferred_positions, args.preferred_anchors
 
     if type(original_pos) == "string" then
@@ -1492,7 +1497,14 @@ function placement.next_to(d, args)
 
     attach(d, placement.next_to, args)
 
-    return fix_new_geometry(ngeo, args, true), pref_name, dir
+    local ret = fix_new_geometry(ngeo, args, true)
+
+    -- Make sure the geometry didn't change, it would indicate an
+    -- "off by border" issue.
+    assert((not osize.width) or ret.width == d.width)
+    assert((not osize.height) or ret.height == d.height)
+
+    return ret, pref_name, dir
 end
 
 --- Restore the geometry.