You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
rig.py: the report cannot tell a helper's head coming apart from a whole one #273
As someone watching the Yard, I want a helper's head never to open into slits or show its inner wall while it bobs and looks around, and as someone changing rig.py, I want the report to fail a helper that does, so a regression cannot ship on the strength of a green report.
Observed
The report's apart check, which catches a character's head layers coming apart (#266), runs on characters only. On helpers it reads geometry that is not a head: it scores whole helpers as broken (gnome_familiar_0 32%, goblin_familiar_2 10%, ear interiors; goblin_helper_2 6%, a dome with modelled holes) while their renders are whole, so it is disabled for them (#270).
For fix/head-walls-266 all 28 helpers were checked by eye instead: every clip's worst pose, three views, all whole. That review has to be repeated by hand for any change that touches helper deformation.
Tried
Six candidate detectors, each measured on labelled examples (main's broken bot_2 .146, goblin_helper_1 .047, goblin_helper_2 .054 against the whole helpers above): depth tolerance past the rest hit; rays restricted to the Head bone's sphere; rays only at Head/Neck-weighted hits; each ray riding the bone of the surface it sees; hidden = unseen from outside; hidden = enclosed and unseen. None separates the broken from the whole: the best keeps main's hero_5/staff_3/staff_4 failing but scores main's broken bot_2 (.0135) below a whole helper (goblin_helper_1 .0115). The cause on the worst false positive is rays aimed at back surface the Head bone does not move, sweeping off it; on the ears, their own thin outer surface is read as enclosed.
To decide
Whether automated detection on helpers is worth building (likely a different measure than fixed rays: e.g. tracking the rest-time neighbours of each head vertex and flagging pairs that part, which is how agent-guild-steam's premium_layers.py works), or whether a visual review of helpers stays the gate for rig.py changes.
Evidence and the candidate scans: D:\agent-guild-media\yard-270-evidence (LOG.md "Layers on helpers", layerscan.py, helpers/).
Activity