Use single-threaded zstd compressor

CHANGELOG.md

@@ -1,6 +1,10 @@
 # Changelog
 ## [Unreleased]
 
+### Fixed
+
+- zstd compression uses a fresh `zstandard` compressor/decompressor per call instead of shared module-level singletons, fixing process segfaults when ingester workers compress documents concurrently (Python < 3.14).
+
 ## [0.61.0] - 2026-06-23
 
 ### Added

haiku_rag_slim/haiku/rag/store/compression.py

@@ -17,7 +17,7 @@ except ImportError:
     # _prepare_document_from_docling), so construct a fresh instance per call
     # rather than sharing a module-level singleton.
     def _zstd_compress(data: bytes) -> bytes:
-        return ZstdCompressor(threads=-1).compress(data)
+        return ZstdCompressor().compress(data)
 
     def _zstd_decompress(data: bytes) -> bytes:
         content_size = get_frame_parameters(data).content_size