fix(openai-compat): unload the outgoing engine on explicit-ID switches

Review follow-up (Greptile/CodeRabbit on #1614): caching instances without
a switch rule would let each distinct explicit engine ID stay resident,
accumulating sidecars / multi-GB in-process models. Mirror
get_active_tts_backend's MM2-01 switch rule: a different explicit ID
(omnivoice included, which resolves to the active engine) unloads the
outgoing instance first, best-effort.

backend/api/routers/openai_compat.py

@@ -158,6 +158,25 @@ _OPENAI_VOICE_ALIASES = {
 # ── TTS: POST /v1/audio/speech ──────────────────────────────────────────────
 
 
+#: Last engine explicitly requested via a `model` ID on this route, and its
+#: instance — mirrors get_active_tts_backend's switch rule (MM2-01): loading a
+#: different explicit engine unloads the outgoing one first, so cached explicit
+#: IDs can't accumulate multi-GB in-process models / sidecars.
+_explicit_engine: dict = {}
+
+
+def _unload_explicit_engine() -> None:
+    inst = _explicit_engine.get("instance")
+    if inst is None:
+        return
+    try:
+        inst.unload()
+    except Exception as exc:  # noqa: BLE001 — a bad unload must not block a switch
+        logger.warning("explicit engine switch: %s.unload() raised: %s",
+                       type(inst).__name__, exc)
+    _explicit_engine.clear()
+
+
 def _resolve_engine(model_id: str):
     """Map an OpenAI model name to a VoiceStudio backend."""
     from services.tts_backend import (
@@ -179,11 +198,20 @@ def _resolve_engine(model_id: str):
             )
         from services.tts_backend import OmniVoiceBackend
         if cls is OmniVoiceBackend:
+            # OmniVoice only ever runs as the shared active engine — the
+            # explicit-omnivoice request is the active-engine request.
+            _unload_explicit_engine()
             return get_active_tts_backend()
         # Cached singleton, not a fresh cls(): SubprocessBackend engines would
         # spawn a sidecar process and reload their model on EVERY request, and
         # register a new atexit hook each time (get_engine_instance's contract).
-        return get_engine_instance_for(model_id)
+        # And on a switch between explicit IDs, unload the outgoing engine so
+        # the cache can't accumulate residents (same rule as the active path).
+        if _explicit_engine.get("id") != model_id:
+            _unload_explicit_engine()
+            _explicit_engine["id"] = model_id
+            _explicit_engine["instance"] = get_engine_instance_for(model_id)
+        return _explicit_engine["instance"]
     except ValueError:
         raise HTTPException(
             status_code=400,

tests/test_openai_speech_engine_cache.py

@@ -62,3 +62,39 @@ def test_unknown_engine_id_still_400s(oc, monkeypatch):
         oc._resolve_engine("not-an-engine")
     assert exc.value.status_code == 400
     assert "Unknown model" in exc.value.detail
+
+
+def test_switching_explicit_engine_ids_unloads_the_outgoing_one(oc, monkeypatch):
+    """Cache must not accumulate residents: a different explicit `model` ID
+    unloads the outgoing engine first (same switch rule as the active-engine
+    path, MM2-01)."""
+    svc = _tts_mod()
+    unloaded = []
+
+    def _make(name):
+        class _B:
+            ident = name
+
+            def __init__(self):
+                pass
+
+            @staticmethod
+            def is_available():
+                return True, "ok"
+
+            def unload(self):
+                unloaded.append(name)
+
+        return _B
+
+    engines = {"a-engine": _make("a"), "b-engine": _make("b")}
+    monkeypatch.setattr(svc, "get_backend_class", lambda i: engines[i])
+
+    first = oc._resolve_engine("a-engine")
+    second = oc._resolve_engine("a-engine")
+    assert first is second
+    assert unloaded == []
+
+    third = oc._resolve_engine("b-engine")
+    assert third is not first
+    assert unloaded == ["a"]