feat(gl): OpenGL 4.1 and the ludic.render3d renderer

`Gl.*` binds the whole OpenGL 4.1 core API — every entry point of the
platform gl3.h with every GL_* constant, generated by `ludic-dev glgen`
with per-call ABI thunks. Windowed builds get an NSOpenGLContext on the
existing window at Retina resolution; headless builds render into an
offscreen CGL context, so a program that uses Gl.* renders and
screenshots identically under the test harness. It links gl.ll, the
thunks and OpenGL.framework only when used; every other build stays
byte-identical.

packages/ludic.render3d is a physically based renderer written on that
surface: HDRI image-based lighting, GPU-generated terrain with scanned
PBR materials, CDLOD, cascaded shadows, glTF with skinning, instanced
vegetation with impostors, procedural grass, water, SSAO, and an HDR
pipeline with bloom, auto-exposure and ACES.

It also carries this session's work on it: the terrain at half its cost
(10.3 -> 5.4 ms of frame), the streaming hitch that got worse the longer
you played, a resize that emptied the world, and the packaging that lets
a game use the renderer from its own repository — `ludic assets`, the
material manifest shipping with the package, and shader lookup falling
back to the install root. See changes/ for each, with its numbers.

The camping game that drove all of it has moved out to its own
repository, Maroon Lake; examples/rendering/smooth.ludic stays as the
renderer's example here.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

changes/terrain-perf.md

@@ -0,0 +1,54 @@
+bump: patch
+type: perf
+**The ground costs half what it did.** Measured in the Maroon Lake game, the terrain was 10.3 ms of
+a 22.2 ms frame; it is now 5.4 ms of 16.4 ms — 45 fps to 61 fps at 1080p, with the
+frame otherwise unchanged (every viewpoint tested stays above 54 dB PSNR against the
+old renderer, with no channel differing by more than 7/255).
+
+**Measure by frame time, not by the pass timers.** `R3D_PROF`'s per-pass
+`GL_TIME_ELAPSED` queries cannot be trusted on this driver: with the ground's shading
+work removed the terrain query fell from 10.5 ms to 1.3 ms while the frame time did
+not move at all. Every number above and below is a median real frame time, taken by
+switching one thing off (`prof_ft_report`); the pass timers are still printed, and are
+still useful for spotting a pass that appears out of nowhere, but they cannot size one.
+`R3D_NOTERRAIN` skips the ground, `R3D_TNEARONLY` / `R3D_TFARONLY` draw every patch
+with one tier's program, and `R3D_RES=<w>x<h>` renders at another size — the three
+switches that say whether a cost is the ground, which tier it is in, and whether it is
+pixels at all.
+
+**Each detail tier is its own program.** `terrain.frag` holds a detailed near tier and
+a cheap far one and chose between them per pixel, so every pixel of the valley walls
+was compiled — and scheduled — for a near path it never ran. CDLOD selection now knows
+which tiers a patch can contain: one that never comes within the split draws with
+`FAR_ONLY`, one wholly inside it with `NEAR_ONLY`, and only the ring of patches that
+straddle the band needs the program that holds both and cross-fades. Pixel-identical,
+and it makes the tiers separately measurable: the near tier costs 7.5 ms over a whole
+frame, the far tier 2.3 ms.
+
+**The sun visibility is its own pass** (`tersun.frag`). The same CDLOD patches are
+rasterised once into a screen-sized R8 buffer that holds nothing but each ground
+pixel's sun visibility, and `terrain.frag` fetches it by fragment coordinate. The pass
+costs 0.27 ms, shares the frame's depth buffer so it doubles as a depth prepass, and
+takes the cascade read out of the shader that covers the screen. It picks its tier —
+filtered PCF near, a single tap far — over the same cross-faded band the ground uses,
+so the boundary is not a contour you can find on the hillside.
+
+**Nothing is sampled for a weight of zero.** The ground sampled all four of its
+materials for every pixel and then blended three of them at zero: a meadow pixel took
+nine taps of triplanar rock, a cliff pixel nine taps of stochastic grass, and every
+pixel in the valley took the snow tile and the four noise fields behind the lake's
+shore wash — a wash that is a hairline along one shore within 120 m of the camera. The
+survey photograph's classification now runs first, because it is what decides which
+materials are present; each material block sits behind its own weight; the ridge field
+that ragged the snow line is skipped 160 m below it, where it cannot change anything;
+and inside the stochastic blend a cell's rotation, offset and rotated gradients are
+computed inside its own test, so a cell whose sharpened weight rounds away costs
+nothing. All of it exact where the weight is zero, and it is most of the win.
+
+Material sampling is what remains (2.8 ms of the 5.4): scanned 2K tiles taken at 16x
+anisotropy on ground seen at a grazing angle. `R3D_ANISO=<n>` sets the filter (the
+default is unchanged at 16; 4 is worth 1.0 ms and 1 is worth 1.7 ms).
+
+The shadow pass is 1.2 ms of the frame and has nothing to give: re-using the far
+cascades between frames — their windows are snapped to a 14 m and a 64 m grid — is
+worth 0.15 ms standing still and nothing while walking, so it is not in the tree.