diff --git a/CHANGELOG.md b/CHANGELOG.md
index 828e25e..b963c52 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,23 @@
 
 Version numbers follow the pattern: `<spaCy-version>-<release-number>`
 
+## [Unreleased]
+
+### Fixed
+- Worker-pool cascade crash on corpora containing a token whose surface form
+  (FORM) is whitespace (e.g. a lone space). Such a token collapsed the
+  tab-separated CoNLL-U line under `str.split()`, leaving 9 fields instead of
+  10 and raising `IndexError` on the score column. The exception was uncaught
+  and killed the whole streaming process, so korapxmltool's worker pool saw a
+  broken pipe, re-queued the in-flight documents, and re-crashed on the same
+  poisoned input until every worker died — observed as running threads dropping
+  off one-by-one after tens of thousands of texts. `CoNLLUP_Token` now parses
+  on the tab delimiter (preserving whitespace FORM columns), falls back to
+  whitespace splitting for space-delimited input, and pads short lines to the
+  10 CoNLL-U columns. As defense-in-depth, a parse/annotation failure for a
+  single document is now logged and skipped while still echoing its
+  `# eot`/`# eof` marker, so one bad document can no longer take down the run.
+
 ## [3.8.11-2] - 2026-06-10
 
 ### Fixed
diff --git a/lib/CoNLL_Annotation.py b/lib/CoNLL_Annotation.py
index 1a8ddf0..06ea564 100644
--- a/lib/CoNLL_Annotation.py
+++ b/lib/CoNLL_Annotation.py
@@ -73,7 +73,20 @@
 
 class CoNLLUP_Token():
     def __init__(self, raw_line, word_ix):
-        info = raw_line.split()
+        # CoNLL-U is tab-separated with 10 columns. Split on the tab so that a
+        # token whose FORM is whitespace (e.g. a surface form that is a single
+        # space) keeps its column position instead of collapsing it -- a bare
+        # raw_line.split() would drop the empty FORM field, shift every column
+        # left, and raise IndexError on info[9], which crashes the whole
+        # streaming process and cascades into broken pipes across the worker
+        # pool. Fall back to a generic whitespace split for space-delimited
+        # inputs, then pad to 10 columns so a genuinely malformed/short line
+        # degrades gracefully rather than killing the stream.
+        info = raw_line.rstrip("\n").split("\t")
+        if len(info) < 10:
+            info = raw_line.split()
+        if len(info) < 10:
+            info = info + ["_"] * (10 - len(info))
         # print(info)
         # [ID, FORM, LEMMA, UPOS, XPOS, FEATS, HEAD, DEPREL, DEPS, MISC]
         # [11, Prügel, Prügel, NN, NN, _, _, _,	_, 1.000000]
diff --git a/systems/parse_spacy_pipe.py b/systems/parse_spacy_pipe.py
index 0432709..33f47de 100644
--- a/systems/parse_spacy_pipe.py
+++ b/systems/parse_spacy_pipe.py
@@ -367,10 +367,19 @@
 	# still processed in CHUNK_SIZE-bounded blocks and streamed out.
 	last_log_time = start
 	for block_lines, terminator in iter_documents(stdin, CHUNK_SIZE):
-		annos, _ = read_conll(iter(block_lines), 0, token_class=token_class, comment_str=args.comment_str, our_foundry="spacy")
-		if annos:
-			dependency_warnings += annotate(annos, total_processed_sents)
-			total_processed_sents += len(annos)
+		# Never let one malformed document take down the whole stream: a parse
+		# or annotation failure here would otherwise crash the process, break
+		# the pipe, and make korapxmltool's worker pool re-queue and re-crash on
+		# the same poisoned input until every worker dies. Log it, skip the
+		# document, and still echo the protocol marker below so the worker pool
+		# releases its in-flight slot.
+		try:
+			annos, _ = read_conll(iter(block_lines), 0, token_class=token_class, comment_str=args.comment_str, our_foundry="spacy")
+			if annos:
+				dependency_warnings += annotate(annos, total_processed_sents)
+				total_processed_sents += len(annos)
+		except Exception as doc_error:
+			logger.error(f"Skipping document after failure ({terminator or 'chunk'}): {str(doc_error)}")
 
 		# Echo the protocol marker back so korapxmltool can pair the output with
 		# the source document and release the in-flight buffer slot, then flush