fix: keep the instance when a turn restores an actor

restore_state rebuilt the actor after a rejected or failed turn without
the instance id, so the activation kept draining with an actor that read
every committed reminder as absent. The activation caches that actor, so
one failed turn poisoned every later read in the same pass.

The three construction sites were the cause, so there is now one.
Activation holds the actor id and the instance id from the start, and
both build_actor and restore_state go through new_actor. A fourth site
cannot forget what it never passes.

Author: cardmagic

lib/solid_objects/activation.rb

@@ -3,6 +3,8 @@
 module SolidObjects
   class Activation
     # @rbs @lease: Lease
+    # @rbs @actor_id: String
+    # @rbs @instance_id: Integer
     # @rbs @actor_class: Class
     # @rbs @actor: Actor
     # @rbs @last_used_at: Float
@@ -17,6 +19,8 @@ def initialize(lease:)
         Instance.find(lease.instance_id)
       end
       @actor_class = SolidObjects.registry.fetch(instance.actor_type)
+      @actor_id = instance.actor_id
+      @instance_id = instance.id
       @actor = build_actor(instance)
       @last_used_at = monotonic_now
       @pass_exhausted = false
@@ -86,10 +90,7 @@ def yield_ready_messages
 
     # @rbs (Hash[String, untyped]) -> void
     def restore_state(state_data)
-      @actor = actor_class.new(
-        actor_id: actor.actor_id,
-        state: State.new(actor_class.definition.state_definition, state_data)
-      )
+      @actor = new_actor(state_data)
     end
 
     # @rbs () -> void
@@ -156,10 +157,15 @@ def build_actor(instance)
       ) do
         actor_class.definition.migrate_state(instance.state_version, instance.state)
       end
+      new_actor(state_data)
+    end
+
+    # @rbs (Hash[String, untyped]) -> Actor
+    def new_actor(state_data)
       actor_class.new(
-        actor_id: instance.actor_id,
+        actor_id: @actor_id,
         state: State.new(actor_class.definition.state_definition, state_data),
-        instance_id: instance.id
+        instance_id: @instance_id
       )
     end
 

sig/generated/lib/solid_objects/activation.rbs

@@ -10,6 +10,10 @@ module SolidObjects
 
     @actor_class: Class
 
+    @instance_id: Integer
+
+    @actor_id: String
+
     @lease: Lease
 
     attr_reader lease: untyped
@@ -59,6 +63,9 @@ module SolidObjects
     # @rbs (Instance) -> Actor
     def build_actor: (Instance) -> Actor
 
+    # @rbs (Hash[String, untyped]) -> Actor
+    def new_actor: (Hash[String, untyped]) -> Actor
+
     # @rbs () -> void
     def release_lease: () -> void
 

test/integration/reminder_cancellation_test.rb

@@ -100,6 +100,14 @@ def arm
       schedule(at: Time.utc(2030, 1, 1)).ping
     end
 
+    def fail_turn
+      raise "turn failed"
+    end
+
+    def record_seen
+      self.seen_on_activate = reminder(:ping)&.name
+    end
+
     def ping
     end
   end
@@ -327,6 +335,20 @@ def state_of(actor_type)
     assert_equal "ping", snapshot.observable_values.fetch("armed")
   end
 
+  test "an actor restored after a failed turn still reads its schedule" do
+    SolidObjects.configuration.max_attempts = 1
+    reference = HookActor.ref("one")
+    reference.async.arm
+    drain
+    # Both messages run in one worker pass, so the activation that the failure
+    # rebuilt is the one that serves the read.
+    reference.async.fail_turn
+    reference.async.record_seen
+    drain
+
+    assert_equal "ping", state_of("cancel-hooks").fetch("seen_on_activate")
+  end
+
   test "a cancel that lands on a claimed occurrence does not fail the scheduler" do
     reference = TrialActor.ref("alice")
     reference.async.start_recurring