Operator-reported via a Bitcoin-films transcript: "Banking on Bitcoin" was
captioned with *The Rise and Rise of Bitcoin*'s description, "Cryptopia" with
*The Bitcoin Standard*'s, and the section header "Documentaries:" bled into the
first card of each group. Read as the model talking nonsense; the model's prose
was correct throughout and only the pairing was wrong.
Several patterns anchor with `(?:^|\n)` so they fire only at a line start.
That makes m.index point at the NEWLINE — one character before the line the
match is really on — so extractDescriptionForTag's window, which walks back
from `matchIndex - 1`, landed on the PREVIOUS line. The description became
"previous line + this item's own text".
Normalised inside the helper rather than at each of its nine call sites, so a
pattern that gains a `(?:^|\n)` anchor later cannot silently reintroduce it.
Fault-injected to prove the tests are not vacuous: with the fix removed, two
fail with exactly the reported strings — 'Documentaries: – Early documentary
fo…' and 'The Rise and Rise of Bitcoin – Early …'. 26/26 with it restored.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
