feat(arr): make the OpenAI endpoint a live setting

The base URL and model move into a cell the backend re-reads per request,
so an operator can repoint it without a restart. Migration 0028 adds the
two columns; DESIGN.md §15 calls both database rows. Construction never
depends on the API key — llama.cpp serves without one.

crates/arr-subs/tests/openai.rs

@@ -16,7 +16,7 @@ use std::time::Duration;
 
 use arr_core::Language;
 use arr_subs::translate::{translate, Backend, Error};
-use arr_subs::{OpenAi, OpenAiConfig};
+use arr_subs::{OpenAi, OpenAiConfig, OpenAiEndpoint};
 use wiremock::matchers::{header, method, path};
 use wiremock::{Mock, MockServer, ResponseTemplate};
 
@@ -319,9 +319,60 @@ async fn an_unsupported_target_is_refused_before_any_request() {
     assert!(matches!(error, Error::UnsupportedTarget { .. }));
 }
 
+/// #220: an edit of `openai_base_url` / `openai_model` reaches the running
+/// backend without a restart — the next batch goes to the new endpoint and
+/// names the new model, with no reconstruction in between.
+#[tokio::test]
+async fn repointing_the_endpoint_moves_the_next_batch() {
+    let first = MockServer::start().await;
+    let second = MockServer::start().await;
+    mount_reply(
+        &first,
+        ResponseTemplate::new(200).set_body_string(chat_response(
+            r#""[{\"number\": 1, \"text\": \"do primeiro\"}]""#,
+        )),
+    )
+    .await;
+    mount_reply(
+        &second,
+        ResponseTemplate::new(200).set_body_string(chat_response(
+            r#""[{\"number\": 1, \"text\": \"do segundo\"}]""#,
+        )),
+    )
+    .await;
+
+    let endpoint = OpenAiEndpoint::new(Some(&format!("{}/v1/", first.uri())), Some("first-model"))
+        .expect("endpoint resolves");
+    let backend =
+        OpenAi::with_endpoint(OpenAiConfig { api_key: None }, endpoint.clone()).expect("builds");
+
+    let cues = vec![arr_subs::translate::BatchCue {
+        number: 1,
+        text: "from the first".to_owned(),
+    }];
+    let batch = arr_subs::translate::Batch {
+        source: Language::Other("en".to_owned()),
+        target: Language::PortuguesePortugal,
+        cues: cues.clone(),
+    };
+    let out = backend.translate(&batch).await.expect("first batch");
+    assert_eq!(out[0].text, "do primeiro");
+
+    endpoint
+        .set(Some(&format!("{}/v1/", second.uri())), Some("second-model"))
+        .expect("repoint");
+    let out = backend.translate(&batch).await.expect("second batch");
+    assert_eq!(out[0].text, "do segundo");
+
+    let sent = second.received_requests().await.expect("requests recorded");
+    let body: serde_json::Value =
+        serde_json::from_slice(&sent.last().expect("one request").body).expect("json body");
+    assert_eq!(body["model"], "second-model");
+}
+
 #[test]
 fn the_backend_answers_to_openai() {
-    let id = OpenAi::new("gpt-4o-mini", OpenAiConfig { api_key: None })
+    let id = OpenAi::new(OpenAiConfig { api_key: None })
         .expect("client builds")
         .id();
     assert_eq!(id.as_str(), "openai");