[cyberspace-cs]

Summary

Fixes #554.

ReferenceRemoverRefiner built the cleaned text with ref_pattern.subn() and cite_pattern.subn(), but then appended the original item instead of the cleaned one:

refined_text, ref_count = ref_pattern.subn('', refined_text)
refined_text, cite_count = cite_pattern.subn('', refined_text)
...
refined_data.append(item)   # item is still the original text

So the operator never removed anything — <ref>...</ref> tags and {{cite ...}} templates stayed in the output column.

In addition, the modified-row counter was incremented twice per changed row (once inside the pattern check and again via a redundant modified flag), so the logged "Total modified items" was double the real count.

Fix

• Append refined_text (the cleaned text) instead of the original item.
• Remove the redundant second increment so each modified row is counted exactly once.

Behavior

For the reproducer in #554:

Input	Before	After
Paris is the capital of France.<ref>Smith 2020</ref>	unchanged	Paris is the capital of France.
`Water boils at 100 C.{{cite web	url=...}}`	unchanged

Summary log for two modified rows: Total modified items: 4 → Total modified items: 2.

Tests

Added test/cpu_only/test_reference_remover_refiner.py (cpu-only):

• removes closed <ref> tags and {{cite}} templates
• removes unclosed forms (matched by the existing patterns)
• leaves rows without markers untouched
• asserts each modified row is counted exactly once via the summary log

pytest -m cpu test/cpu_only/test_reference_remover_refiner.py → 4 passed.

[cyberspace-cs]

Closing this as a duplicate — I missed #555 when I opened it. Same root cause, same fix (reassign item before appending, count each modified row once), and #555 was opened five days earlier, so it should be the one that lands.

My branch additionally covers the double-counted "Total modified items" log line and a no-substitution case; I can send those as a small follow-up to #555 if you want them. Sorry for the noise.