Add benchmark-driven model promotion workflow and pipeline stages

tests/test_aman.py

@@ -111,11 +111,21 @@ class FakeUnsupportedLanguageModel:
 class FakeAIProcessor:
     def __init__(self):
         self.last_kwargs = {}
+        self.warmup_calls = []
+        self.warmup_error = None
+        self.process_error = None
 
     def process(self, text, lang="auto", **_kwargs):
+        if self.process_error is not None:
+            raise self.process_error
         self.last_kwargs = {"lang": lang, **_kwargs}
         return text
 
+    def warmup(self, profile="default"):
+        self.warmup_calls.append(profile)
+        if self.warmup_error:
+            raise self.warmup_error
+
 
 class FakeAudio:
     def __init__(self, size: int):
@@ -212,6 +222,32 @@ class DaemonTests(unittest.TestCase):
 
         self.assertEqual(desktop.inject_calls, [("good morning Marta", "clipboard", False)])
 
+    @patch("aman.stop_audio_recording", return_value=FakeAudio(8))
+    @patch("aman.start_audio_recording", return_value=(object(), object()))
+    def test_editor_failure_aborts_output_injection(self, _start_mock, _stop_mock):
+        desktop = FakeDesktop()
+        model = FakeModel(text="hello world")
+        ai_processor = FakeAIProcessor()
+        ai_processor.process_error = RuntimeError("editor boom")
+
+        daemon = self._build_daemon(
+            desktop,
+            model,
+            verbose=False,
+            ai_processor=ai_processor,
+        )
+        daemon._start_stop_worker = (
+            lambda stream, record, trigger, process_audio: daemon._stop_and_process(
+                stream, record, trigger, process_audio
+            )
+        )
+
+        daemon.toggle()
+        daemon.toggle()
+
+        self.assertEqual(desktop.inject_calls, [])
+        self.assertEqual(daemon.get_state(), aman.State.IDLE)
+
     def test_transcribe_skips_hints_when_model_does_not_support_them(self):
         desktop = FakeDesktop()
         model = FakeModel(text="hello")
@@ -242,7 +278,7 @@ class DaemonTests(unittest.TestCase):
         self.assertEqual(used_lang, "auto")
         self.assertIn("Docker", model.last_kwargs["hotwords"])
         self.assertIn("Systemd", model.last_kwargs["hotwords"])
-        self.assertIn("Preferred vocabulary", model.last_kwargs["initial_prompt"])
+        self.assertIsNone(model.last_kwargs["initial_prompt"])
 
     def test_transcribe_uses_configured_language_hint(self):
         desktop = FakeDesktop()
@@ -300,7 +336,7 @@ class DaemonTests(unittest.TestCase):
         daemon_verbose = self._build_daemon(desktop, FakeModel(), cfg=cfg, verbose=True)
         self.assertTrue(daemon_verbose.log_transcript)
 
-    def test_ai_processor_is_initialized_during_daemon_init(self):
+    def test_editor_stage_is_initialized_during_daemon_init(self):
         desktop = FakeDesktop()
         with patch("aman._build_whisper_model", return_value=FakeModel()), patch(
             "aman.LlamaProcessor", return_value=FakeAIProcessor()
@@ -308,7 +344,47 @@ class DaemonTests(unittest.TestCase):
             daemon = aman.Daemon(self._config(), desktop, verbose=True)
 
         processor_cls.assert_called_once_with(verbose=True, model_path=None)
-        self.assertIsNotNone(daemon.ai_processor)
+        self.assertIsNotNone(daemon.editor_stage)
+
+    def test_editor_stage_is_warmed_up_during_daemon_init(self):
+        desktop = FakeDesktop()
+        ai_processor = FakeAIProcessor()
+        with patch("aman._build_whisper_model", return_value=FakeModel()), patch(
+            "aman.LlamaProcessor", return_value=ai_processor
+        ):
+            daemon = aman.Daemon(self._config(), desktop, verbose=False)
+
+        self.assertIs(daemon.editor_stage._processor, ai_processor)
+        self.assertEqual(ai_processor.warmup_calls, ["default"])
+
+    def test_editor_stage_warmup_failure_is_fatal_with_strict_startup(self):
+        desktop = FakeDesktop()
+        cfg = self._config()
+        cfg.advanced.strict_startup = True
+        ai_processor = FakeAIProcessor()
+        ai_processor.warmup_error = RuntimeError("warmup boom")
+        with patch("aman._build_whisper_model", return_value=FakeModel()), patch(
+            "aman.LlamaProcessor", return_value=ai_processor
+        ):
+            with self.assertRaisesRegex(RuntimeError, "editor stage warmup failed"):
+                aman.Daemon(cfg, desktop, verbose=False)
+
+    def test_editor_stage_warmup_failure_is_non_fatal_without_strict_startup(self):
+        desktop = FakeDesktop()
+        cfg = self._config()
+        cfg.advanced.strict_startup = False
+        ai_processor = FakeAIProcessor()
+        ai_processor.warmup_error = RuntimeError("warmup boom")
+        with patch("aman._build_whisper_model", return_value=FakeModel()), patch(
+            "aman.LlamaProcessor", return_value=ai_processor
+        ):
+            with self.assertLogs(level="WARNING") as logs:
+                daemon = aman.Daemon(cfg, desktop, verbose=False)
+
+        self.assertIs(daemon.editor_stage._processor, ai_processor)
+        self.assertTrue(
+            any("continuing because advanced.strict_startup=false" in line for line in logs.output)
+        )
 
     @patch("aman.stop_audio_recording", return_value=FakeAudio(8))
     @patch("aman.start_audio_recording", return_value=(object(), object()))