Consumer-side helpers for Sagents.Publisher producers.

Captures the boilerplate of:

1. Subscribing to a producer by id rather than pid.
2. Monitoring the producer so we know when it dies.
3. Watching Phoenix.Presence arrivals for that id (to handle crash-restart with a new pid, or Horde migration to a new node).
4. Re-subscribing automatically when the producer reappears.
5. Cleaning up on caller exit (best-effort — the producer's monitor on the caller will clean us up anyway).

Targets two consumer shapes:

• Plain GenServer / process — call subscribe_to_agent/2 and friends directly; pair with handle_publisher_down/3 and handle_presence_diff/3 from your handle_info/2.
• LiveView (or anything with a socket) — use Sagents.Subscriber injects handle_info/2 clauses for :DOWN and Phoenix presence diffs that delegate to those helpers.

Subscription handle

Each active subscription is tracked in the caller's local subs map (kept on the LiveView socket under socket.private.sagents_subs, or threaded through manually for plain-process callers).

Map shape:

%{
  {:agent, agent_id} => %{
    channel: :main | :debug,
    server_pid: pid() | nil,
    monitor_ref: reference() | nil,
    state: :subscribed | :pending
  },
  {:filesystem, scope_key} => %{...}
}

States:

• :subscribed — we have a live subscription on server_pid, monitored.
• :pending — we want to subscribe but the producer isn't running. Will retry on the next presence arrival.

Departure vs arrival

Monitors are reliable for departure (:DOWN fires within a scheduler tick of process death, even cross-node when the connection drops). Phoenix.Presence is reliable for arrival (presence_diff with joins). We never depend on presence_diff.leaves — if it's delayed, we keep the :subscribed state until :DOWN fires, which it will.