feat: duration_ms-logging, bulk-semaphore und erweitertes README

- Pipeline-Stages (download/extract/embed/qdrant) loggen jetzt duration_ms
- bulk-import dispatcht mit Semaphore(4) statt unbounded → Backpressure
- README dokumentiert Webhook-Payload-Schema mit curl-Beispiel
- README enthaelt Recovery-Runbook (dim-mismatch, crash-recovery, single-file reindex)

app/bulk.py

@@ -1,3 +1,4 @@
+import asyncio
 import logging
 import xml.etree.ElementTree as ET
 from pathlib import PurePosixPath
@@ -18,6 +19,17 @@ logger = logging.getLogger(__name__)
 router = APIRouter()
 
 
+# Limits parallel pipeline execution during a bulk import so we don't
+# saturate Ollama/WebDAV with thousands of concurrent requests.
+BULK_CONCURRENCY = 4
+_bulk_semaphore = asyncio.Semaphore(BULK_CONCURRENCY)
+
+
+async def _process_with_semaphore(file_path: str, event_type: EventType) -> None:
+    async with _bulk_semaphore:
+        await process_file(file_path, event_type)
+
+
 class BulkRequest(BaseModel):
     path: str = Field(min_length=1)
 
@@ -93,7 +105,7 @@ async def bulk_import(
         ext = PurePosixPath(f).suffix.lstrip(".").lower()
         if ext not in SUPPORTED_TYPES:
             continue
-        background.add_task(process_file, f, EventType.CREATED)
+        background.add_task(_process_with_semaphore, f, EventType.CREATED)
         dispatched += 1
 
     logger.info(