Bug: response_end span in granite33 uses full response length instead of sentence length

[psschwei]

This was flagged as worthy of followup when reviewing #818 :

In granite33/output.py:226 — The line computes:

mellea/mellea/formatters/granite/granite3/granite33/output.py

Line 226 in 301ca3e

citation["response_end"] = index + len(response_text_without_citations)

• response_text is the individual sentence associated with this citation (looked up from response_sents_by_citation_id)
• response_text_without_citations is the entire response with all citation tags stripped
• index is where the individual sentence starts within the full response

So response_end ends up being (start of sentence) + (length of entire response), which will always overshoot past the end of the actual sentence — potentially past the end of the string entirely.

The correct formula should be:

citation["response_end"] = index + len(response_text)

The granite32 version does this correctly:

mellea/mellea/formatters/granite/granite3/granite32/output.py

Lines 291 to 293 in 301ca3e

	citation["response_end"] = last_response_text_match["begin_idx"] + len(
	response_text
	)

Impact: Every citation span in granite 3.3 output will have a response_end that points to roughly the end of the full response rather than the end of the cited sentence. Downstream consumers that slice response[begin:end] would get a much larger substring than intended.

[planetf1]

Fix reference for agents:

mellea/formatters/granite/granite3/granite33/output.py:227 — change:

citation["response_end"] = index + len(response_text_without_citations)

to:

citation["response_end"] = index + len(response_text)

The granite32 equivalent at granite3/granite32/output.py:292-293 is correct and can be used as reference. Add a regression test in test/formatters/granite/test_granite33_output.py.