LGTM, but let's not risk interrupting work on the release, so merge this after branch-off.
If that goes well, I'd support a backport when it's been on nixos-unstable for about a month.

After discussing this with @hsjobeki in a call today, we discovered that this PR currently doesn't work well for @hsjobeki's usecase, because while the freeform definitions would be exposed, the priorities of individual attributes would have to be recalculated outside the module evaluation. This would notably also be a problem for standard non-freeform options with nested priorities.

Instead, the valueMeta direction from #391544 allows this PR to work better because each merge (no matter how deep it occurs), can expose more detailed merge results, which could then be used for the freeform type merge as well, like this:

diff --git a/lib/modules.nix b/lib/modules.nix
index 4691018c68a0..611f6e46aac9 100644
--- a/lib/modules.nix
+++ b/lib/modules.nix
@@ -226,6 +226,10 @@ let
             '';
           };
 
+          _module.freeformMerged = mkOption {
+            internal = true;
+            visible = false;
+            readOnly = true;
+            type = types.raw;
+          };
+
           _module.specialArgs = mkOption {
             readOnly = true;
             internal = true;
@@ -242,6 +246,8 @@ let
             moduleType = type;
           };
           _module.specialArgs = specialArgs;
+
+          _module.freeformMerged = freeformMerged;
         };
       };
 
@@ -266,21 +272,26 @@ let
 
       options = merged.matchedOptions;
 
+      # For definitions that have an associated option
+      declaredConfig = mapAttrsRecursiveCond (v: !isOption v) (_: v: v.value) options;
+
+      freeformDefs = map (def: {
+        file = def.file;
+        value = setAttrByPath def.prefix def.value;
+      }) merged.unmatchedDefns;
+
+      freeformMerged = declaredConfig._module.freeformType.merge.v2 {
+        loc = prefix;
+        defs = freeformDefs;
+      };
+
       config =
         let
 
-          # For definitions that have an associated option
-          declaredConfig = mapAttrsRecursiveCond (v: !isOption v) (_: v: v.value) options;
 
           # If freeformType is set, this is for definitions that don't have an associated option
           freeformConfig =
-            let
-              defs = map (def: {
-                file = def.file;
-                value = setAttrByPath def.prefix def.value;
-              }) merged.unmatchedDefns;
-            in
-            if defs == [ ] then { } else declaredConfig._module.freeformType.merge prefix defs;
+            # Probably can be simplified with freeformMerged.isDefined or so?
+            if freeformDefs == [ ] then { } else freeformMerged.value;
 
         in
         if declaredConfig._module.freeformType == null then
diff --git a/lib/types.nix b/lib/types.nix
index 35171b28c799..07fa9d03e88b 100644
--- a/lib/types.nix
+++ b/lib/types.nix
@@ -852,6 +852,7 @@ let
                     v.optionalValue.value
                 ) evals;
                 valueMeta.attrs = mapAttrs (n: v: v.checkedAndMerged.valueMeta) evals;
+                valueMeta.evals = evals;
               };
           };
 

                 valueMeta.attrs = mapAttrs (n: v: v.checkedAndMerged.valueMeta) evals;
+                valueMeta.evals = evals;

This looks redundant or wrong.
What's the purpose of this?

+      freeformDefs = map (def: {

Floating this stuff out causes a space leak.
Although we have those all over the place, it's worth checking the impact on the final heap size.
Considering that this particular binding is only referenced by config (indirectly, but very much so), it'd be better to inline it, so that it can be released when freeformMerged has consumed it. (Best if .isDefined works, as commented)
Otherwise it will still be referenced through the scope that contains it, which is in turn likely referenced by any of the many thunks produced by evalModules.

• Naive thunk closures cause unnecessary space leaks nix#8285

+      freeformDefs = map (def: {

Floating this stuff out causes a space leak. Although we have those all over the place, it's worth checking the impact on the final heap size. Considering that this particular binding is only referenced by config (indirectly, but very much so), it'd be better to inline it, so that it can be released when freeformMerged has consumed it. (Best if .isDefined works, as commented) Otherwise it will still be referenced through the scope that contains it, which is in turn likely referenced by any of the many thunks produced by evalModules.

* [Naive thunk closures cause unnecessary space leaks nix#8285](https://github.com/NixOS/nix/issues/8285)

Thanks, thats good to know. i'll try move it back into the let binding again after we are done with #391544

We've basically discovered, that it would be better to integrate with merge.v2 because that would give us the new merge interface that natively allows exposing metadata