End a line at a five-em gap, and rejoin a title split across lines

A per-field audit (166 of 177 fields byte-identical) surfaced a
regression the corpus average had been hiding: gasimova's title fell
from 1.000 to 0.706 back at 1bc9ef1, while its overall rField still rose
because Background gained 0.396. _flatten_top_band puts everything in the
banner on common baselines so the byline's markers can find their names,
but that also drops a logo set in the top corner onto the title's line
("... A Universal Framework FAIR DATA") and a poster-ID badge onto 42's
("... Carcinogens #LB108").

Sharing a baseline does not make two pieces of text one line. Word
spacing runs about a quarter of an em; the logo and the badge stand five
ems clear (335pt at 63pt type, 490pt at 96pt), and nothing else in any
banner exceeds 0.8em. A gap of LINE_MAX_GAP ems now ends the run. It is
applied after the superscript merge, since a marker row is legitimately
spread across the width of its byline and must not be torn apart. Swept:
4.0 keeps the logo and badge out while leaving 4446908 intact, which 2.5
split.

A title wrapped across lines then arrives as one block per line, so
rejoin the header blocks set in the largest type. That was aimed at
gasimova and instead fixed 4446908, whose wrapped title went 0.706 to
1.000 - a general pattern, not a special case.

Also tried and reverted: classifying any large top-zone text as a title
fragment, to reach gasimova's second title line (which a narrow logo
excludes by dragging col_start above it). It cost 4519718 0.372 on its
banner and AISec2025 0.401, for one field's gain. Not worth it; the
approach is in the history if someone wants it.

Corpus rField 0.735 to 0.737 (0.729 at session start), w 0.976, rGlobal
0.835, 20/21, 12/13 numbered, 13/13 logic-OK, suite green (260 passed).
Every affiliation string in the corpus is unchanged and verified. Eleven
fields improved, one still down: gasimova's title, the standing cost of
the flatten, which buys that poster its byline (+0.154), Background
(+0.396) and correct affiliations for all seven authors. Worth it, but
recorded rather than averaged away.

Author: jimnoneill