Add 90-min timeout kill to PropagationGridWorker

lib/microwaveprop/workers/propagation_grid_worker.ex

@@ -5,10 +5,12 @@ defmodule Microwaveprop.Workers.PropagationGridWorker do
   Fetches f00-f18 to provide 18-hour forecast timeline.
   """
 
+  # `unique.period` matches the 3-hour cron so a retrying job can't be
+  # duplicated by the next cron firing.
   use Oban.Worker,
     queue: :propagation,
     max_attempts: 3,
-    unique: [period: 3600, states: [:available, :scheduled, :executing, :retryable]]
+    unique: [period: 10_800, states: [:available, :scheduled, :executing, :retryable]]
 
   alias Microwaveprop.Propagation
   alias Microwaveprop.Propagation.BandConfig
@@ -22,6 +24,16 @@ defmodule Microwaveprop.Workers.PropagationGridWorker do
 
   require Logger
 
+  # Hard ceiling per run. A healthy run is ~60 min (19 forecast hours × ~3 min
+  # each). 90 min gives 1.5× headroom and is under the 3-hour cron cadence, so
+  # a stuck wgrib2/HTTP call can't block the queue indefinitely. Oban kills
+  # the executing process on timeout, which closes linked ports and cascades
+  # SIGKILL to any child wgrib2 subprocess.
+  @run_timeout_ms 90 * 60 * 1000
+
+  @impl Oban.Worker
+  def timeout(_job), do: @run_timeout_ms
+
   @max_forecast_hour 18
 
   @impl Oban.Worker