Fix three VMAF calc state management bugs

1. Resolve 'will retry' failures on success: When CRF search retry
   succeeds and transitions to crf_searched, resolve any unresolved
   CRF search failures for that video. Previously 24+ encoded videos
   had stale unresolved failure records.

2. Reset state when sync deletes VMAFs: When video_upsert detects
   metadata changes and deletes VMAFs, also reset crf_searched videos
   back to analyzed. Previously sync could nuke VMAFs via
   delete_vmafs_for_video without fixing the state, leaving 486+
   invalid crf_searched videos with no VMAF data.

3. Run reset_crf_searched_without_vmaf periodically: Schedule the
   orphan reset every 5 minutes instead of only on encoder startup.
   Acts as a safety net for any future state inconsistencies.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: mjc

lib/reencodarr/ab_av1/crf_search.ex

@@ -547,6 +547,7 @@ defmodule Reencodarr.AbAv1.CrfSearch do
   defp transition_to_crf_searched(video) do
     case Retry.retry_on_db_busy(fn -> Media.mark_as_crf_searched(video) end) do
       {:ok, _} ->
+        Media.resolve_crf_search_failures(video.id)
         :ok
 
       {:error, reason} ->

lib/reencodarr/encoder/broadway/producer.ex

@@ -12,6 +12,8 @@ defmodule Reencodarr.Encoder.Broadway.Producer do
 
   # Poll every 2 seconds to check for new work
   @poll_interval_ms 2000
+  # Reset orphaned/invalid states every 5 minutes
+  @orphan_reset_interval_ms 300_000
 
   # After 900 consecutive unavailable polls (~30 minutes), attempt recovery
   @recovery_threshold 900
@@ -31,6 +33,7 @@ defmodule Reencodarr.Encoder.Broadway.Producer do
   def init(_opts) do
     send(self(), :reset_orphaned)
     schedule_poll()
+    schedule_orphan_reset()
     {:producer, %{pending_demand: 0, consecutive_unavailable: 0}}
   end
 
@@ -51,6 +54,7 @@ defmodule Reencodarr.Encoder.Broadway.Producer do
   def handle_info(:reset_orphaned, state) do
     Media.reset_orphaned_encoding()
     Media.reset_crf_searched_without_vmaf()
+    schedule_orphan_reset()
     {:noreply, [], state}
   end
 
@@ -93,6 +97,10 @@ defmodule Reencodarr.Encoder.Broadway.Producer do
     Process.send_after(self(), :poll, @poll_interval_ms)
   end
 
+  defp schedule_orphan_reset do
+    Process.send_after(self(), :reset_orphaned, @orphan_reset_interval_ms)
+  end
+
   # Public for testing
   @doc false
   # :available means encoder responded and is free — reset counter

lib/reencodarr/media/video_upsert.ex

@@ -147,11 +147,28 @@ defmodule Reencodarr.Media.VideoUpsert do
     if not being_marked_encoded and
          VideoValidator.should_delete_vmafs?(existing_video, new_values) do
       delete_vmafs_for_video(existing_video.id)
+      reset_crf_searched_state(existing_video.id)
     end
 
     :ok
   end
 
+  # When VMAFs are deleted during sync, any crf_searched video must be reset to
+  # analyzed — the old VMAF data is invalid and the video needs re-searching.
+  defp reset_crf_searched_state(video_id) do
+    {count, _} =
+      from(v in Video,
+        where: v.id == ^video_id and v.state == :crf_searched
+      )
+      |> Repo.update_all(
+        set: [state: :analyzed, chosen_vmaf_id: nil, updated_at: DateTime.utc_now()]
+      )
+
+    if count > 0 do
+      Logger.info("Reset crf_searched video #{video_id} → analyzed after VMAF deletion")
+    end
+  end
+
   @spec handle_bitrate_preservation(
           %{String.t() => any()},
           map() | nil,