Make the panel tile grid a parameter

[rasros]

Why

The panel cuts every picture into a GRID x GRID lattice, embeds each cell, and averages the 25 cell distances into one number per member. That average is what makes small features invisible: a change confined to one cell is divided by 25.

Measured on duck-v14's eye, which the run drew with its polarity inverted (white disc, dark crescent, against the target's black pupil with a white rim highlight). Hand-correcting the two fills and rescoring:

grid	cell	v14 as-is	eye corrected	delta
1	700	0.157864	0.126567	−0.03130
2	350	0.159684	0.150497	−0.00919
3	233	0.257846	0.263664	+0.00582
5	140	0.296499	0.299647	+0.00315
8	87	0.351762	0.364387	+0.01263

At the current grid of 5 the eye's bounding box is 11.8% of one 140px cell and straddles a cell boundary, so it is roughly 0.5% of the score surface — which is where the per-member deltas land, at ±0.0005 on distances of 0.036 to 0.096. The three members are near-tied and a 2–1 split decides it.

Caveat this does not settle: both drawings are wrong in different ways, so "which wrong is closer" is a judgement call, and this is one hand-edited pair on one target. It is a reason to be able to vary the grid, not yet a reason to change its default.

What

GRID was a module constant read by _tiles. It is now a parameter:

• _tiles(image, grid), where grid <= 1 returns the picture itself
• EnsembleScorer(grid=...), threaded through get_scorer, the runner, and main
• --panel-grid N, defaulting to the existing 5

No behaviour changes unless the flag is passed.

Tests

Two added: a grid of 1 hands the members the whole picture and 5 gives 25 cells; a grid's tiles cover the picture without overlap. Full suite green at 642.

[rasros]

Superseded: rather than make the grid configurable, the tiling is being removed outright. See the follow-up PR.