A thread without DiscussionForumPosting markup, where every poster writes
once, still passed the teaser test: each post opens with a linked name that
no other post repeats. On katzen-forum.net with the JSON-LD removed and one
post per poster, recall took the whole container (3,573 chars, seven user
cards instead of one).
Teasers hold neither a description list nor a quote, while forum posts keep
the poster's user card in one and quote each other, so an entry holding
either now rules the container out.
Thread pages line up their posts as sibling <article> elements too, so in
recall mode the container took every post along with its user card and
quotes (katzen-forum.net, XenForo).
The container is now used only when every entry is a teaser: its first
heading links to a page of its own, which no other entry links to. Poster
names come back along a thread or carry no link. On top of that it is never
used on pages already detected as thread forums (DiscussionForumPosting).
On correctiv.org the article page preloads the next twelve full stories as
sibling <article> elements, so with favor_recall the listing container took
all of them (98k chars instead of 9.6k). A teaser strip beside a short body
(mopo.de) was also taken over the body.
The container is now used only when the element BODY_XPATH picks on its own
is one of its entries and shorter than the recall sure-thing length
(10x min_extracted_size).
A blog index, archive or category page holds several <article> siblings.
BODY_XPATH's "(.//article)[1]" keeps the first entry only, so the page
looks under-extracted and escalates to justext, which emits paragraphs
and drops every heading.
Give the recall pass a listing expression that takes the container of
the sibling articles, so the retry returns all entries with their heads
and wins the escalation. Recall only: on an article page the same shape
is a related-posts strip, whose container would hide the body.
dualfroz
upstream