Turning on doubles via scons precision=double will change the docs

[fire]

Godot version

60d0317

System information

Windows 11, Nvidia 3000

Issue description

Turning on doubles will change the doctool behaviour.

Problems:

1. The API should not change. Means you can't use the same API dump for multiple targets
2. Add a CI check for this, we already need to compile twice

Steps to reproduce

1. build godot with scons precsion=double
2. run godot with --doctool
3. check git diff

Minimal reproduction project

N/A

[akien-mga]

Can you attach the full diff?

[fire]

Sure. I'm not at a ready state so I'll get it in a few days.

[akien-mga]

Here's the diff (from #71575):

diff --git a/doc/classes/Curve3D.xml b/doc/classes/Curve3D.xml
index 362d792..96311c7 100644
--- a/doc/classes/Curve3D.xml
+++ b/doc/classes/Curve3D.xml
@@ -40,7 +40,7 @@
 			</description>
 		</method>
 		<method name="get_baked_tilts" qualifiers="const">
-			<return type="PackedFloat32Array" />
+			<return type="PackedFloat64Array" />
 			<description>
 				Returns the cache of tilts as a [PackedFloat32Array].
 			</description>
diff --git a/doc/classes/HeightMapShape3D.xml b/doc/classes/HeightMapShape3D.xml
index 206981e..0243bb0 100644
--- a/doc/classes/HeightMapShape3D.xml
+++ b/doc/classes/HeightMapShape3D.xml
@@ -10,7 +10,7 @@
 	<tutorials>
 	</tutorials>
 	<members>
-		<member name="map_data" type="PackedFloat32Array" setter="set_map_data" getter="get_map_data" default="PackedFloat32Array(0, 0, 0, 0)">
+		<member name="map_data" type="PackedFloat64Array" setter="set_map_data" getter="get_map_data" default="PackedFloat64Array(0, 0, 0, 0)">
 			Height map data, pool array must be of [member map_width] * [member map_depth] size.
 		</member>
 		<member name="map_depth" type="int" setter="set_map_depth" getter="get_map_depth" default="2">
diff --git a/doc/classes/PhysicsDirectSpaceState2D.xml b/doc/classes/PhysicsDirectSpaceState2D.xml
index d4cb073..10b43cc 100644
--- a/doc/classes/PhysicsDirectSpaceState2D.xml
+++ b/doc/classes/PhysicsDirectSpaceState2D.xml
@@ -12,7 +12,7 @@
 	</tutorials>
 	<methods>
 		<method name="cast_motion">
-			<return type="PackedFloat32Array" />
+			<return type="PackedFloat64Array" />
 			<param index="0" name="parameters" type="PhysicsShapeQueryParameters2D" />
 			<description>
 				Checks how far a [Shape2D] can move without colliding. All the parameters for the query, including the shape and the motion, are supplied through a [PhysicsShapeQueryParameters2D] object.
diff --git a/doc/classes/PhysicsDirectSpaceState2DExtension.xml b/doc/classes/PhysicsDirectSpaceState2DExtension.xml
index fbbb98a..8f19f4c 100644
--- a/doc/classes/PhysicsDirectSpaceState2DExtension.xml
+++ b/doc/classes/PhysicsDirectSpaceState2DExtension.xml
@@ -16,8 +16,8 @@
 			<param index="4" name="collision_mask" type="int" />
 			<param index="5" name="collide_with_bodies" type="bool" />
 			<param index="6" name="collide_with_areas" type="bool" />
-			<param index="7" name="closest_safe" type="float*" />
-			<param index="8" name="closest_unsafe" type="float*" />
+			<param index="7" name="closest_safe" type="double*" />
+			<param index="8" name="closest_unsafe" type="double*" />
 			<description>
 			</description>
 		</method>
diff --git a/doc/classes/PhysicsDirectSpaceState3D.xml b/doc/classes/PhysicsDirectSpaceState3D.xml
index cc1cf8a..78829e2 100644
--- a/doc/classes/PhysicsDirectSpaceState3D.xml
+++ b/doc/classes/PhysicsDirectSpaceState3D.xml
@@ -12,7 +12,7 @@
 	</tutorials>
 	<methods>
 		<method name="cast_motion">
-			<return type="PackedFloat32Array" />
+			<return type="PackedFloat64Array" />
 			<param index="0" name="parameters" type="PhysicsShapeQueryParameters3D" />
 			<description>
 				Checks how far a [Shape3D] can move without colliding. All the parameters for the query, including the shape, are supplied through a [PhysicsShapeQueryParameters3D] object.
diff --git a/doc/classes/PhysicsDirectSpaceState3DExtension.xml b/doc/classes/PhysicsDirectSpaceState3DExtension.xml
index 4297846..036c5cd 100644
--- a/doc/classes/PhysicsDirectSpaceState3DExtension.xml
+++ b/doc/classes/PhysicsDirectSpaceState3DExtension.xml
@@ -16,8 +16,8 @@
 			<param index="4" name="collision_mask" type="int" />
 			<param index="5" name="collide_with_bodies" type="bool" />
 			<param index="6" name="collide_with_areas" type="bool" />
-			<param index="7" name="closest_safe" type="float*" />
-			<param index="8" name="closest_unsafe" type="float*" />
+			<param index="7" name="closest_safe" type="double*" />
+			<param index="8" name="closest_unsafe" type="double*" />
 			<param index="9" name="info" type="PhysicsServer3DExtensionShapeRestInfo*" />
 			<description>
 			</description>

[akien-mga]

Fixing this will require breaking compatibility, but it's a bit late to do for 4.0 anyway, so lets work on that for 4.1.

[dsnopek]

I did a little bit of research and this one is trickier than I initially thought.

The switch to using real_t in both those files was on purpose, in these two PRs (both by @aaronfranke):

• Improve Curve with real_t and naming #53868
• Use real_t in physics code #45519

I didn't look as closely at the physics one, but looking at Curve3D, it does make sense for it to use double for its own math if Vector3 is using double (as it does with precision=double).

So, if we decide that exposed APIs changing between float and double is a bad thing we want to avoid... Does that mean these should always use double and PackedFloat64Array, wasting some space in precision=single? Or, that we should convert to float and PackedFloat32Array when exposed to the API, losing some precision when using precision=double?

Or, is the API changing when doing precision=double just part of the cost of having this as an option?

[aaronfranke]

I don't see it as a big deal that the docs and APIs change between precision levels, I think it's fully expected that the different precision levels are not compatible.

However, if we did pick one single size for exposing something, it should be double, not single. We already have the precedent that GDScript Variant float and Array[float] are always double precision.

[fire]

I also support picking the widened type.