feat(csg): -texchart, what each texture costs the bsp

[metita]

Why

Measured both of the maps in Documents/Mapping, compiled with their real flags:

lump	zm_azteca	zm_eichen
texdata	2.263.180 B — 82,8%	1.842.804 B — 69,9%
lightdata	236.490 B — 8,7%	405.981 B — 15,4%
visdata	15.210 B — 0,6%	12.444 B — 0,5%
entdata	9.700 B — 0,4%	17.617 B — 0,7%
total	2.732.216 B	2.637.540 B

Textures are the map. -chart prints one texdata total and stops, so there is no way to tell which textures that is or whether the resolution is doing any work.

This also kills two ideas that looked reasonable before measuring: lightmap deduplication targets 9-15% of the file and would recover a fraction of that, and duplicate textures are essentially absent (0.1% on eichen, 0 on azteca). Neither is worth building.

What it does

-texchart, off by default, after WriteMiptex:

     bytes  share   size       painted    oversampled  texture
    174892   7.7%   512x256        83520        0.5x  zg2026jun3
     87852   3.9%   256x256      2683422        0.0x  BT_wall04
     87852   3.9%   256x256        38776        1.5x  BT_door01

• bytes come from the gaps between consecutive lump entries, so padding is included rather than recomputed from the WAD3 layout.
• painted is the world surface area the texture ends up on, accumulated per texinfo in WriteFace for hull 0 only.
• oversampled is texture pixels over pixels actually displayed. The texture axes are scaled by texels-per-unit, so their lengths turn world area into displayed pixels exactly. Above 4x, halving each axis still leaves a pixel per pixel on screen.
• Textures with byte-identical first mips are flagged.

Why the area cross-reference matters

A naive "halve every texture 256px and up" projection claimed 70% off zm_azteca. With the area accounted for, the real answer is that nothing on azteca is oversampled — those textures tile across very large surfaces, and downscaling would only look worse. The naive number was measuring nothing.

Actual results:

• zm_azteca: 32 textures, 2.269.304 B, nothing carrying 4x more pixels than it displays.
• zm_eichen: 90 textures, 1.941.720 B, 3 worth halving (118.326 B, 6.1%), one duplicated pair (black/white, 32x32, 2.172 B).

The duplicate figure cross-checks against an independent Python parse of the same bsp.

Safety

Off by default, and AccumulateTextureArea returns immediately when it is off. It only reads back the lump CSG has already written, so it cannot change compiler output. All five tools build clean; with the flag absent CSG prints nothing extra.