research/paper-code-links — every number in the write-up

rows: 390  REFUSED: 0  OTHER: 0
(no repository refused this server and none answered anything but
 200 or 404, so the refusal carve-out in the rule never had to fire)

old     n=190 alive=182 gone= 8 rate=0.9579  95% Wilson [0.9191, 0.9785]
recent  n=200 alive=193 gone= 7 rate=0.9650  95% Wilson [0.9295, 0.9829]

old - recent = -0.0071  95% [-0.0454, +0.0312]
Seven to eight years of ageing moves availability by less than one
point and the interval covers zero comfortably. The dead ones are
not rotted: they are the same share dead in a cohort months old.

old     18 of 182 live repositories answered only after a redirect (9.9%)
recent   7 of 193 live repositories answered only after a redirect (3.6%)

If GitHub stopped honouring renames, the 2018 rate would be 0.8632 rather than 0.9579.
The string printed in the paper is wrong in those cases. The
platform is doing the remembering, not the citation.

One recent URL ends in '_', which is the shape of a truncation in the abstract or in my normaliser.
Dropping it: recent n=199 k=193 rate=0.9698 — the gate still passes and no verdict moves.