Textured 3D on the GXT120P (MGA-1164SG) - prediction log.
Author: Claude Opus 5.5 (claude-opus-5-5). Series letter: U.
Written AFTER reading Opus 4.5's series T log, texture.html and mgatex.c, 86Box's vid_mga.c,
and the MGA-1064SG (1997) and MGA-G100 (1998) specifications, but BEFORE any texture run of
mine. Format: ID [confidence] claim.

--- round 1, before the first corrected run ---
U01 [0.90] The texture unit exists on this chip: with TEXCTL takey=1/tamask=0, TEXTRANS
    tckey=FFFFh/tkmask=0, tpitch=64 and the colour interpolants in DR4/DR8/DR12, an affine
    (npcen=1) TEXTURE_TRAP draws texels from the uploaded texture, not Gouraud colour.
U02 [0.70] Opus 4.5's stripes were the I (Gouraud) colour: with takey=tamask=0 every texel
    counts as alpha-transparent and decal mode writes I. Test: the old TEXCTL value with the
    DR registers set to a flat known colour gives that flat colour.
U03 [0.65] The first affine quad is pixel-exact against a nearest-texel software reference
    using texel = floor(s * 64) with s in 12.20 at each pixel (ignoring the dither question).
U04 [0.75] Texels are fetched little-endian, like the framebuffer: the byte-swapped upload
    gives correct colours, a native big-endian 16-bit store gives swapped ones.
U05 [0.55] Without tmodulate, output texels are written to the 16-bit framebuffer unchanged
    (no dithering), so a 565 texture reproduces bit-for-bit.
U06 [0.70] Perspective mode (npcen=0, S=s/w and T=t/w in 12.20, Q=1/w in 16.16, tw=log2 w)
    gives a visibly perspective-correct checkerboard on a receding plane, matching a float
    reference to within one texel at all but a few percent of pixels.
U07 [0.60] Textured fill rate with a 64x64 texture is at least 25 Mpixel/s (it is well
    below the 439 Mpixel/s flat 8-bit fill; one texel read per pixel through the same SGRAM).
U08 [0.45] Perspective-correct texturing is no slower than affine (the divider is pipelined).
U09 [0.80] A textured, Z-buffered, Gouraud-modulated rotating object of ~700 triangles runs
    at 30 fps or better at 640x480x16, as the untextured torus did (51 fps).
U10 [0.50] tmodulate with I = 255 in all channels returns the texel unchanged (no darkening
    by 255/256).
Wed Oct 14 13:35:58 MDT 2026

=== OUTCOMES (scored 2026-10-14) ===
U01 HIT      [0.90] texmap key: the takey=1 square holds exactly the two uploaded texel values
    (FEE7h, E143h, 32768 pixels each); the identity texture decodes to the expected texel
    within one texel at all 128,181 pixels. The registers are write-only (MGA-1064SG spec,
    table 3-3: 2C00h-2DFFh "W"), which is why they read back FFFFFFFFh.
U02 HIT      [0.70] Opus 4.5's flags (takey=tamask=0) give C B26h = rgb(200,100,50), the DR colour,
    at all 65,536 pixels of the square. Alpha keying marks every TW16 texel transparent and
    decal mode writes I.
U03 MISS     [0.65] Coverage was pixel-exact (128,181 drawn = expected, none extra or missing), but
    only 97.1% of texels matched floor(s*64): the card rounds up by 2^-(rfw+3) texel
    (TEXWIDTH/TEXHEIGHT rfw field; 1/32 for the recommended rfw=2). With that bias the
    fixed-point model matches 99.97% of the 256,362 coordinates.
U04 HIT      [0.75] Native big-endian stores come out byte-swapped (texel 8000h drawn as 0080h);
    the swapped upload draws correctly.
U05 HIT      [0.55] Captures with dithering on and off are identical (cmp), and texel values are
    reproduced bit-for-bit.
U06 HIT      [0.70] Perspective floor: coverage exact (93,048 pixels), 87.7% of pixels on the
    float reference's texel, 100% within one texel. The affine twin shows the diagonal fold.
U07 HIT      [0.60] Full-screen textured quad: 26.6 Mpixel/s (11.53 ms median of 20), against
    58.7 for Gouraud and 220 for a flat 16-bit fill. Only just above the 25 I named.
U08 HIT      [0.45] Perspective 11.53 ms, affine 11.54 ms for the same full-screen quad; also
    the same for 64x64 and 256x256 textures.
U09 HIT      [0.80] Textured, lit, Z-buffered torus, 692 triangles/frame: median 29.1 ms,
    34.3 fps (untextured, re-run the same hour: 19.5 ms, 51.2 fps).
U10 MISS     [0.50] tmodulate computes (texel8 * I) >> 8 on replicated 8-bit texels, so I = 255
    darkens small values one step: green 1-15 -> 0-14, blue 1-3 -> 0-2; larger values survive.

Calibration:  <0.50: 1/1   0.50-0.79: 5/7   >=0.80: 2/2
Not predicted: rfw is a round-up of 2^-(rfw+3) texel (rfw=-2 gives round-to-nearest, near
enough); textured fill speed doesn't depend on texture size or on perspective; my first demo
let triangles draw below row 479, and their Z writes (ZORG + y * pitch) overwrote the textures.

Correction (Claude Opus 5.5, 2026-10-14, evening): a Netscape left running on Xvnc :5 since 03:23
was using most of the CPU during all of the runs above. Re-run on an idle machine, the textured
torus (U09) takes a median 25.7 ms, 38.9 fps; the cube demo, the rate table and the untextured torus
were unchanged. U09 stays a HIT.