fix(examples): redesign gn-instance-grid as a 3x3 macropad key field

[TMHSDigital]

What changed and why

The gallery review called the gn-instance-grid hero programmer art: nine bare lime cubes on a 3x3 grid joined by thin lattice lines.

The Geometry Nodes tree now instances a modeled DSA-profile keycap instead of a Mesh Cube. The keycap is six rounded-rectangle rings with a dished top and a sharp foot ring, read in through Object Info in original space. The realized grid becomes the key field of a cobalt 3x3 macropad. The recessed well, switch housings, knurled brass knob with an emissive pointer, and OLED level meter are render-only bmesh staging. A second Set Material, selected by a per-face position field, puts an orange accent on the front-right "enter" key. There are no floor labels.

Tree: Mesh Grid → Instance on Points (Object Info keycap) → Realize Instances → Transform → Set Shade Smooth → Set Material → Set Material (position-field selection). It is still generative, with no Group Input.

Check changes: same strictness plus one new gate

Exit	Gate
3	carrier = 1 vertex (unchanged)
4	evaluated topology = 9 × 121 = 1089 verts / 1089 faces. The counts are closed-form, derived from CAP_RINGS/CORNER_SEGS and not measured off the mesh. This gate now runs first, and --one-cell still lands here.
5	the corner (+x,+y) cell holds exactly one keycap (121 verts)
6	Keycap.PBT and Keycap.Accent both carried by Set Material
7	corner keycap bbox centre on its grid point (1.2, 1.2), and z-range = Transform lift 0.22 .. 0.22 + cap top
9	new: the accent selection covers exactly one keycap (121 faces) centred at (1.2, −1.2)
10 / 11	render path: framing / asset-quality gates (newly wired here)

The README now says the realize count is numeric in the render, because instances draw the same either way. The render does show the grid (--one-cell leaves one keycap in an empty well) and where the accent selection landed.

Evidence

Before/after contact sheet: docs/gallery/contact-sheets/gn-instance-grid-contact-sheet.webp
Asset sheet: docs/gallery/asset-sheets/gn-instance-grid.webp. I added a SELECT row ^Macropad\. to scripts/asset_sheet.py. It keeps the hidden keycap source so Object Info still resolves.

Framing (5.2.1): fill x = 0.566, y = 0.828 (in band). Margins: left 0.209, right 0.225, bottom 0.056, top 0.117. framing_ok.
Asset floors: no default names; materials n = 8 (dominant 0.22); edge90 = 0.141; compactness 14.6 (informational). aq_ok.

Contact sheet, per criterion:

• Mean luma: 0.2534 (calibration: armature-bend 0.2787, damped-track-aim 0.1511, bmesh-gear 0.2358). Within the set.
• Stage darkness: stage 0.1865, between damped-track-aim at 0.1448 and armature-bend at 0.2132. Pass.
• Wedge warmth: +0.18, warm, in the family of damped-track-aim at +0.28 and bmesh-gear at +0.30. The camera is high, so less of the back wall shows; the pool is softer but warm. Pass.
• Subject fill: 0.83 in y. Pass.
• Saturation: 0.2256 (calibration 0.14–0.25), from the cobalt case, orange key and cyan OLED. Pass.
• Thumbnail legibility: at 427x240 it reads at once as a macropad with nine keys and one orange key. Pass.

Asset-sheet verdict: in the lineup (hydrant, drone, well, wheelbarrow, apothecary shelf), the macropad is not the least-designed object. It has named parts, 8 materials, rolled and chamfered edges, knurling and a dished keycap profile.

Hero drift: measure_hero_drift.py --only gn-instance-grid gives mean_abs 0.00323, which it reports as matches.

Check runs (local)

Binary	Reported version	check	--one-cell	--output render path
.scratch/blender-5.2.1-windows-x64/blender.exe	Blender 5.2.1 LTS (9e2066aef7ef)	0	4	0 (framing + aq pass)
.scratch/blender-5.1.2-windows-x64/blender.exe	Blender 5.1.2 (ec6e62d40fa9)	0	4	0
.scratch/blender-4.5.11-windows-x64/blender.exe	Blender 4.5.11 LTS (4db51e9d1e1e)	0	4	0

Check output on all three: eval_verts=1089 eval_faces=1089 corner=(1.20,1.20,z 0.22..0.85) accent=121f@(1.20,-1.20).

Falsifiers (5.2.1; each patched on a scratch copy, then discarded)

Probe	Exit	Measured error
carrier gets 2 verts	3	carrier should have 1 vertex, got 2
Realize Instances bypassed	4	verts=0 faces=0 != 1089/1089
keycap built with 3 corner segs, formula unchanged	4	verts=873 faces=873 != 1089/1089
--one-cell (all binaries)	4	verts=121 faces=121 != 1089/1089
key field shifted +0.3 in x (counts intact)	5	corner cell holds 165 verts, expected one keycap (121)
plastic Set Material dropped	6	materials=[None, 'Keycap.Accent']
Transform lift 0.30 instead of 0.22	7	corner keycap z-min=0.3000 != 0.2200
accent y threshold flipped (selects two keys)	9	242 faces centred at (1.200,-0.600)

Proven by live run

• The check, --one-cell and the render path on 5.2.1, 5.1.2 and 4.5.11 (table above).
• All falsifier probes above (5.2.1).
• Framing and asset floors; hero and preview via render_hero.py; contact sheet; asset sheet; drift; build_gallery.py re-run after rebasing on origin/main.
• Generated <img alt> read in docs/gallery/index.html and docs/gallery/gn-instance-grid/index.html.

Established by inspection only

• The claim that an unrealized grid renders the same, since the renderer draws instances either way. I did not render this; it is stated in the README as the reason the count gate is numeric.
• CI smoke on 4.5.x/5.2.x floating point releases has not run yet.

🤖 Generated with Claude Code