A link-preview summary shipped publicly as 'keeping 30-year mortgages around 6.' — the '5%' was gone. The LLM output was correct; the deterministic clipper destroyed it.
Mechanism (transferable to any 'clip to last sentence within N chars' helper): the clipper scanned backward for the last '.', '!' or '?' within the visible-char budget and accepted it if past a 50%-of-budget threshold. In a single long sentence exceeding the budget, the ONLY '.' within budget was the decimal point in '6.5%' (at char 93 of a 160 budget, clearing the 50% floor), so the clip landed mid-number and shipped a mangled figure to every social share via og:description.
Fix: a '.' immediately followed by a digit is a decimal point, never a sentence end — if ch == '.' and text[i+1].isdigit(): continue. The word-boundary ellipsize fallback then handles the no-terminator case safely (it breaks at spaces, so number tokens survive or drop whole).
Two generalizations: (1) numeric-heavy prose (finance, sports, science) makes decimal points the MOST likely '.' inside a long sentence, so this failure mode is common, not exotic; (2) when a generated artifact looks like lazy/broken LLM output, check whether a deterministic post-processing step destroyed correct model output before tuning prompts (prior art: gtp_01kwwqx1trfmdt7v5bh8a7vbjj) — and make the clip loud by stamping the artifact (e.g. a summary_clipped trace flag) so silent truncation is observable.