From 381a344435dbddc5b2cd3c7c955f5e9f7d30e4e3 Mon Sep 17 00:00:00 2001
From: Min Huang <70873102+min0625@users.noreply.github.com>
Date: Wed, 8 Jul 2026 07:56:21 +0000
Subject: [PATCH] docs: warn about repetition-loop risk in long-input manual
 test

Repeated-phrase input used in case 15 can send some real models
(gpt-4o-mini, llama3.1:8b) into a runaway repetition loop that looks
like a hang; recommend varied content (testdata/sample.txt) instead
when testing against a real model.
---
 docs/manual-testing.md | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/docs/manual-testing.md b/docs/manual-testing.md
index 9492ad7..086240f 100644
--- a/docs/manual-testing.md
+++ b/docs/manual-testing.md
@@ -295,6 +295,14 @@ mint -t zh-TW -v < /tmp/long.txt
 # ... three translated paragraphs, separated by blank lines
 ```
 
+> **Against a real model** (not a mock), prefer a varied-content document over the
+> repeated-phrase one above — e.g. `testdata/sample.txt` duplicated once
+> (`cat testdata/sample.txt testdata/sample.txt`). A single phrase repeated dozens of
+> times is degenerate input that can send some models (observed: `gpt-4o-mini`,
+> `llama3.1:8b`) into a runaway repetition loop, producing a chunk many times longer
+> than the input and taking minutes instead of seconds — a model quirk, not a chunking
+> bug, but it can look like a hang during manual testing.
+
 In rotation mode, language detection samples the first chunk that contains letters
 (falling back to the first chunk if none do), so detection cost does not grow with
 document length. If that sample comes back language-neutral, a single-chunk document is