Simplify editor cleanup and keep live ASR metadata

Keep the daemon path on the full ASR result so word timings and detected language survive into the editor pipeline instead of falling back to a plain transcript string.

Add PipelineEngine.run_asr_result(), have aman call it when live ASR data is available, and cover the word-aware alignment behavior in the daemon tests.

Collapse the llama cleanup flow to a single JSON-shaped completion while leaving the legacy pass1/pass2 parameters in place as compatibility no-ops.

Validated with PYTHONPATH=src python3 -m unittest tests.test_aiprocess tests.test_aman.

src/aman.py

@@ -142,6 +142,7 @@ def _process_transcript_pipeline(
     stt_lang: str,
     pipeline: PipelineEngine,
     suppress_ai_errors: bool,
+    asr_result: AsrResult | None = None,
     asr_ms: float = 0.0,
     verbose: bool = False,
 ) -> tuple[str, TranscriptProcessTimings]:
@@ -161,7 +162,10 @@ def _process_transcript_pipeline(
             total_ms=asr_ms,
         )
     try:
-        result = pipeline.run_transcript(processed, language=stt_lang)
+        if asr_result is not None:
+            result = pipeline.run_asr_result(asr_result)
+        else:
+            result = pipeline.run_transcript(processed, language=stt_lang)
     except Exception as exc:
         if suppress_ai_errors:
             logging.error("editor stage failed: %s", exc)
@@ -546,6 +550,7 @@ class Daemon:
                     stt_lang=stt_lang,
                     pipeline=self.pipeline,
                     suppress_ai_errors=False,
+                    asr_result=asr_result,
                     asr_ms=asr_result.latency_ms,
                     verbose=self.log_transcript,
                 )