Row counts in v2 were wrong (2,599 -> 2,078 events; 1,403 -> 1,087 usable
labels). Every row now carries its excerpt (v2 had 1,094 NULLs), and the
seed false-positive warning is replaced by a measured per-row column:
the substring-fabrication class that produced a `Y` event from "Howdy
y'all!" is now empty, 0 of 405 usable seed rows. Adds
is_new_content_contradicted, which flags the 23% of that LLM flag's
positives contradicted by pre-patch price rows. Card carries the
project's own retractions: the 20x cost error, the directional null, the
tautological fill rate, and that no execution regime is profitable.
