[ii-impy]

Closes #164.

Not ready to merge until the rig test passes. The real depth-camera test on the Ultra 9 is the sign-off. The checklist is below.

Summary

There are two parts: a player change and a new Setup.

The player gains an "image" channel source (ADR-0023, which amends ADR-0021). Each load takes the next image from the Setup's images/ folder, in shuffled order:

 Rotation._load(name)
   setup = load_setup(dir / name)            # + setup.images: <setup>/images/*.png|jpg|jpeg
+  image = _next_image(setup)                # ShuffleBag per Setup; corrupt files dropped
+    decode_image(path, cover=window,        # EXIF-upright, row 0 = bottom,
+                 max_side=GL_MAX_TEXTURE)   # downscaled only while it still covers the window
   renderer.load(setup, image=image)         # unbound → reads 0, iChannelResolution 0

The new Setup, shaders/watercolor_smudge/, shows a watercolor painting. The thin, moving parts of a visitor's silhouette smear it, and it heals once the room empties:

B  original   image cover-fit, or a procedural watercolor sunset when images/ is empty
C  motion     Body softened + eased (10 Hz → smooth), last frame, thinness, empty timer
D  canvas     paint pushed along the moving silhouette edge (normal flow)
                × thinness (limbs and head yes, torso no)
                × balance gate (appearing, vanishing or flickering Body doesn't count)
              → finger streaks, pigment-style mixing, wetness
              → heals toward B: 2 s after the room empties, ~95 % in 10 s
Image         the bound image sampled directly where untouched; canvas + paper grain where smeared

All tuning values are named constants at the top of common.glsl. The design is in docs/superpowers/specs/2026-11-07-watercolor-smudge-design.md, with the approved playable sketch beside it. The build plan is in docs/superpowers/plans/2026-11-07-watercolor-smudge.md.

Also: shader_app.py can now go fullscreen on a chosen display. Use shader_app.display in captcha-settings.json, or --display N to override it. The window is sized from that display's own desktop, so a 1600×1200 projector beside a 3072×1920 main monitor gets the whole picture. Before this, the window took the main monitor's size and you only saw part of it.

Evidence

• Before: 343 FundingCAPTCHA tests. No "image" source: load_setup raises "unknown source(s) ['image']".
  After: 377 pass (.venv/bin/python -m pytest -q in ShowControl/FundingCAPTCHA/). The new behaviour tests render through real GL:
  • test_still_flickering_person_leaves_paint_alone: a still limb with 1-camera-px rim flicker and edge dropout at 10 Hz → max canvas change < 0.02
  • test_moving_limb_smears_most_of_its_path → over 30 % of the path smeared. It was 7 % before the softening fix that came out of the look check.
  • test_moving_torso_does_not_smear → < 0.03
  • test_heal_waits_for_delay_then_heals_in_heal_time, test_heal_pauses_when_someone_returns
  • test_image_covers_screen_keeping_aspect, test_image_stays_sharp_outside_smear, test_black_paint_at_centre_still_smears
  • test_rotation_shuffles_images_per_load, test_rotation_skips_corrupt_image, test_rotation_with_only_corrupt_images_binds_nothing, test_decode_image_applies_exif_orientation
• Look check, rendered headlessly at 800×450 with a walking, waving stick figure: walking smudges the water, the wave drags the sun into the cloud, and after everyone leaves it heals clean, which matches the approved sketch. A temporary portrait image also checked out: it cover-fits, stays sharp where untouched, smears, and heals. It wasn't committed.

Known limitation, pinned by test_reaching_arm_along_its_axis_does_not_smear_yet:

• What happens: an arm reaching straight out along its own length doesn't smear. It only gains Body at the tip, so the balance gate reads it as appearing. Waving, swinging and walking do smear.
• What was tried: two bounded fixes, each of which broke the torso or flicker rejection.
• What to try at the rig: raise BALANCE_LO/BALANCE_HI if reaching matters more than flicker.

Rig checklist: run python3 shader_app.py --camera --setup watercolor_smudge, and press R to reload after edits.

• Stand still for 30 s: no smear.
• Walk through, wave, and reach.
• Leave and watch the heal.
• Note the fps log line.

To tune, adjust these constants:

Symptom	Constant
Smear too broad	THIN, RING_R1/RING_R2
Too weak	SMEAR_STRENGTH, BODY_SOFT
A still person smears	MIN_FLOW, BODY_EASE_S
Heal speed	HEAL_TIME_S

Merge Danger

Door: two-way. The new code is additive: a new channel source, a new Setup folder, and one line added to setups.json. Reverting removes it cleanly.

Blast Radius: rotation.

• watercolor_smudge joins the shader rotation for every event that runs shader_app.py.
• Rotation._load now decodes images for any Setup that binds "image". No existing Setup does, so their behaviour is unchanged.
• A large image adds a decode hitch at the change to this Setup.

Deferred minors from the final review:

• After a corrupt file is skipped, a just-shown image can repeat.
• A broken edit followed by R still uses up an image.
• A failed image upload can leave the renderer half-loaded.
• The comment on PRESENT_HEAL overstates what it does in 16-bit.
• T_CLAMP is unused.
• A variable named step shadows the GLSL built-in.
• The flicker test would be stricter on a checker image.

🤖 Generated with Claude Code