fix: repair deferred imports to preserve module-level names for test patching (#831)

A previous commit moved SDK imports into __init__/methods and
stashed them on self, which broke @patch targets in 24 unit tests.

This fixes the approach: chunker and pdf_decoder use module-level
sentinels with global/if-None guards so imports are still deferred but
patchable. Google AI Studio reverts to standard module-level imports
since the module is only loaded when communicating with Gemini.
Keeps lazy loading on other imports.

trustgraph-flow/trustgraph/chunking/recursive/chunker.py

@@ -10,6 +10,8 @@ from prometheus_client import Histogram
 from ... schema import TextDocument, Chunk, Metadata, Triples
 from ... base import ChunkingService, ConsumerSpec, ProducerSpec
 
+RecursiveCharacterTextSplitter = None
+
 from ... provenance import (
     chunk_uri as make_chunk_uri, derived_entity_triples,
     set_graph, GRAPH_SOURCE,
@@ -41,8 +43,12 @@ class Processor(ChunkingService):
         self.default_chunk_size = chunk_size
         self.default_chunk_overlap = chunk_overlap
 
-        from langchain_text_splitters import RecursiveCharacterTextSplitter
-        self.RecursiveCharacterTextSplitter = RecursiveCharacterTextSplitter
+        global RecursiveCharacterTextSplitter
+        if RecursiveCharacterTextSplitter is None:
+            from langchain_text_splitters import (
+                RecursiveCharacterTextSplitter as _cls,
+            )
+            RecursiveCharacterTextSplitter = _cls
 
         if not hasattr(__class__, "chunk_metric"):
             __class__.chunk_metric = Histogram(
@@ -52,7 +58,7 @@ class Processor(ChunkingService):
                          2500, 4000, 6400, 10000, 16000]
             )
 
-        self.text_splitter = self.RecursiveCharacterTextSplitter(
+        self.text_splitter = RecursiveCharacterTextSplitter(
             chunk_size=chunk_size,
             chunk_overlap=chunk_overlap,
             length_function=len,
@@ -105,7 +111,7 @@ class Processor(ChunkingService):
             chunk_overlap = int(chunk_overlap)
 
         # Create text splitter with effective parameters
-        text_splitter = self.RecursiveCharacterTextSplitter(
+        text_splitter = RecursiveCharacterTextSplitter(
             chunk_size=chunk_size,
             chunk_overlap=chunk_overlap,
             length_function=len,

trustgraph-flow/trustgraph/chunking/token/chunker.py

@@ -10,6 +10,8 @@ from prometheus_client import Histogram
 from ... schema import TextDocument, Chunk, Metadata, Triples
 from ... base import ChunkingService, ConsumerSpec, ProducerSpec
 
+TokenTextSplitter = None
+
 from ... provenance import (
     chunk_uri as make_chunk_uri, derived_entity_triples,
     set_graph, GRAPH_SOURCE,
@@ -41,8 +43,10 @@ class Processor(ChunkingService):
         self.default_chunk_size = chunk_size
         self.default_chunk_overlap = chunk_overlap
 
-        from langchain_text_splitters import TokenTextSplitter
-        self.TokenTextSplitter = TokenTextSplitter
+        global TokenTextSplitter
+        if TokenTextSplitter is None:
+            from langchain_text_splitters import TokenTextSplitter as _cls
+            TokenTextSplitter = _cls
 
         if not hasattr(__class__, "chunk_metric"):
             __class__.chunk_metric = Histogram(
@@ -52,7 +56,7 @@ class Processor(ChunkingService):
                          2500, 4000, 6400, 10000, 16000]
             )
 
-        self.text_splitter = self.TokenTextSplitter(
+        self.text_splitter = TokenTextSplitter(
             encoding_name="cl100k_base",
             chunk_size=chunk_size,
             chunk_overlap=chunk_overlap,
@@ -104,7 +108,7 @@ class Processor(ChunkingService):
             chunk_overlap = int(chunk_overlap)
 
         # Create text splitter with effective parameters
-        text_splitter = self.TokenTextSplitter(
+        text_splitter = TokenTextSplitter(
             encoding_name="cl100k_base",
             chunk_size=chunk_size,
             chunk_overlap=chunk_overlap,

trustgraph-flow/trustgraph/decoding/pdf/pdf_decoder.py

@@ -15,6 +15,9 @@ from ... schema import Document, TextDocument, Metadata
 from ... schema import librarian_request_queue, librarian_response_queue
 from ... schema import Triples
 from ... base import FlowProcessor, ConsumerSpec, ProducerSpec, LibrarianClient
+
+PyPDFLoader = None
+
 from ... provenance import (
     document_uri, page_uri as make_page_uri, derived_entity_triples,
     set_graph, GRAPH_SOURCE,
@@ -128,7 +131,12 @@ class Processor(FlowProcessor):
                 fp.write(base64.b64decode(v.data))
                 fp.close()
 
-            from langchain_community.document_loaders import PyPDFLoader
+            global PyPDFLoader
+            if PyPDFLoader is None:
+                from langchain_community.document_loaders import (
+                    PyPDFLoader as _cls,
+                )
+                PyPDFLoader = _cls
             loader = PyPDFLoader(temp_path)
             pages = loader.load()