Absorb, reviewed from several angles (#9441)
One finding per angle, read from the code at 47c57bd6. Work that follows is filed as its own issue.
- Security. Reach is sound: a declaration alone does nothing, only a non-agent site admin absorbs, readers alone get a menu entry,
/apps/openre-checks access,pageis a path on the app's own host, and a forge-owned variable name cannot be set. Two gaps: the menu name and page follow the latest production commit rather than the admin's act (#9541), and deleting an app leaves its feature rows and plain-text settings behind (#9542). Settings readable by repository readers, and by previews, is documented and unchanged. - Without the crew. Nothing quietly needs a model: absorb, settings, redeploy and the menu call no model, and
test/apps-features.test.tsasserts no run is started with no credential. The crew cannot absorb either (agent accounts are refused), which is intended. No issue. - Experience. Discovery, confirmation and refused-save behaviour are thin, and the menu mixes features among the forge's own links (#9544).
- Product. Packs cannot absorb, absorbed features show nowhere but the menu, and a forked declaration shows no provenance (#9545).
- Speed. The Features list reads git per app per request, and the menu does a few lookups per absorbed app on every page (#9543). Not measured; no regression at the current two features is claimed.
- Credit. Ben (benjivers) is not a user on this forge and the crew cannot reach his; the pull description carries the findings and the owner passes them on.
Since #9681 the menu entry's title is the declared summary, copied at absorb and on accept; see absorb.md. Features absorbed earlier show no title until re-accepted.