noe/arthur
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Pin every dense track to the feature whose presence it follows

Browse source 
The absence mapping was asserted for two of the ten dense tracks. Each block
hands `pack` an :absent predicate that derives a feature from a track INDEX, so
the predicate and the vector literal beside it agree only by hand, in four
places — and three of the four blocks had nothing checking them.

Give each feature a DIFFERENT gap window, so a track wired to the wrong feature
shows up as absence inside somebody else's window. A single shared window passes
under any permutation, which is the failure port-plan mechanical fact #2 warns
about: swap left for right and every part is still roughly where it belongs, so
it survives inspection.

Verified by mutation, since a test that cannot fail is worth nothing. Swapping
the brow block's two tracks, and swapping eye-block tracks 1 and 3 while leaving
0 and 2 correct, both now fail loudly; neither was caught before.

Also pin the teeth. They are their own feature so they can carry their own
:area :teeth parameters, which means an occluded mouth sets no bit on them; they
are dropped regardless because they stencil on :mouth-in and scene/finish drops a
node whose stencil drew nothing. Both frames come from an unannotated reference
clip, because the interior only draws on an open mouth and a frame the mouth was
shut on would have passed for the wrong reason.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B87NVmiU36qQmN9gmFYnJ9
This commit is contained in:
Olive Vaughn 2026-09-27 23:38:26 -04:00
parent 35ef150b48
commit b6517f837a
1 changed files with 95 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

95
frontend/test/arthur/flow/freeze_test.cljs
View file

@ -402,6 +402,101 @@
(is (contains? drawn-nodes :mouth)))
(is (empty? (scene/problems sc)))))
(deftest each-dense-track-follows-its-own-features-presence
;; Every feature is given a DIFFERENT gap, so a track wired to the wrong one
;; shows up as absence inside somebody else's window. One shared gap would pass
;; under any permutation of the four tracks in the eye block, and that is the
;; failure docs/port-plan.md warns about twice: swap left for right and every
;; part is still roughly where it belongs, so it survives inspection.
;;
;; This is the mapping `pack` cannot check for itself. Each block hands it an
;; :absent predicate that derives a feature from a track INDEX, so the predicate
;; and the vector literal beside it have to agree by hand, in four places.
(let [windows {:eye-r (set (range 10 15))
:eye-l (set (range 20 25))
:brow-r (set (range 30 35))
:brow-l (set (range 40 45))}
presence (into {} (map (fn [[id gap]]
[id (mapv #(not (contains? gap %))
(range take/frames))]))
windows)
c (freeze/clip (assoc take/params :name "per-track-gaps")
(assoc @take/measured :presence presence))
sc (:scene c)
at (fn [id path f]
(ch/value-at (get-in sc [:nodes id :channels path]) f (:store c)))
;; Node, channel, and the feature whose gap it must follow. Every dense
;; track of the eye, iris, brow and brow-position blocks appears once.
tracks [[:eye-r [:geom :pts] :eye-r]
[:eye-r-in [:geom :pts] :eye-r]
[:eye-l [:geom :pts] :eye-l]
[:eye-l-in [:geom :pts] :eye-l]
[:iris-r [:xform :pos] :eye-r]
[:iris-l [:xform :pos] :eye-l]
[:brow-r [:geom :pts] :brow-r]
[:brow-l [:geom :pts] :brow-l]
[:brow-r [:xform :pos] :brow-r]
[:brow-l [:xform :pos] :brow-l]]]
(is (empty? (scene/problems sc)))
(doseq [[id path owner] tracks
[feature gap] windows
f gap]
(if (= feature owner)
(is (ch/nothing? (at id path f))
(str id " " path " is not absent at " f " with " feature " occluded"))
(is (not (ch/nothing? (at id path f)))
(str id " " path " follows " feature "'s gap at frame " f))))))
(deftest teeth-follow-the-mouth-through-their-stencil-and-not-through-a-mask
;; Teeth are their own feature, so that they can carry their own :area :teeth
;; parameters — which means an occluded MOUTH sets no absence bit on them. They
;; do not need one: they are stencilled by :mouth-in, and scene/finish drops a
;; node whose stencil drew nothing. So the coupling is real, and it is the
;; stencil rule rather than the mask that enforces it. The rule itself is
;; asserted in scene-test; what is pinned here is the wiring that relies on it.
(let [gap (set (range 40 60))
teeth-in {;; The mouth's own inner ring standing in for a pixel-derived
;; contour: real geometry and no nils, so every frame has a
;; value that could be dropped.
:contours (:inner @take/measured)
:shown (vec (repeat take/frames true))}
clip (fn [nm extra]
(freeze/clip (assoc take/params :name nm)
(merge (assoc @take/measured :teeth teeth-in) extra)))
ref (clip "teeth-reference" nil)
occ (clip "teeth-mouth-gap"
{:presence {:mouth (mapv #(not (contains? gap %))
(range take/frames))}})
;; Hoisted: the resolver caches its order and reuses its buffers, so the
;; node ids come out before the next frame is asked for.
nodes-at (fn [c]
(let [r (scene/resolver (:scene c) (:store c) pal/index-of)]
(fn [f] (into #{} (map :node) (r f)))))
ref-at (nodes-at ref)
occ-at (nodes-at occ)
;; The interior only draws on an OPEN mouth, so both frames are picked
;; from the UNANNOTATED clip. Same articulation either side of the gap,
;; which is what stops this passing on a frame the mouth was shut anyway.
open (filter #(contains? (ref-at %) :teeth) (range take/frames))
inside (first (filter gap open))
outside (first (remove gap open))]
(is (= :mouth-in (get-in (:scene occ) [:nodes :teeth :stencil]))
"teeth stop inheriting the mouth's absence if this stops being their stencil")
(is (some? inside) "no open-mouth frame inside the gap to test with")
(is (some? outside) "no open-mouth frame outside the gap to test with")
(when (and inside outside)
;; Annotating the MOUTH sets no bit on the teeth: they are a feature of
;; their own and nothing named them.
(is (not (ch/nothing?
(ch/value-at (get-in (:scene occ) [:nodes :teeth :channels [:geom :pts]])
inside (:store occ)))))
;; They are dropped anyway — :mouth-in drew nothing to clip them against.
(is (not (contains? (occ-at inside) :mouth-in)))
(is (not (contains? (occ-at inside) :teeth)))
;; And the same articulation outside the gap still draws them.
(is (contains? (occ-at outside) :teeth)))
(is (empty? (scene/problems (:scene occ))))))
(deftest an-undetected-frame-has-no-pose-at-all
;; A subject that was not on the frame has NO VALUE, which is different from a
;; part being switched off. The mask lands on every block of the freeze, so an