feat(stdlib): add Noise.* — deterministic fixed-point procedural noise (#3)

A Noise.* namespace for procedural generation, implemented entirely in Q16.16
fixed point over an integer permutation hash so a seed reproduces the exact same
field on every platform and run (native/headless/wasm) — the determinism edge
over float noise that drifts across CPUs.

  - value2 / perlin2 / simplex2  — value, gradient, and simplex noise -> [-1,1]
  - fbm2(x,y,seed,octaves)       — fractal Brownian motion (octaves of simplex)
  - cellular2 / cellular2_id     — Worley F1 distance + nearest-cell id
  - unit(n)                      — remap [-1,1] -> [0,1]

Covers issue phases 1–2 fully plus cellular from phase 3; domain warp, ridged/
billow, and sample1/sample3 remain as follow-ups. Pure integer IR, C-free;
cellular/fbm reuse the math prelude's fx_sqrt.

- examples/library/noise.ludic: asserts the invariants a fixed-point generator
  must hold (Perlin == 0 at lattice points, every sampler within [-1,1],
  reproducibility, seed sensitivity, non-negative cellular distance). Wired into
  `x test` (now 52 passed).
- docs: a new Noise section + per-symbol pages; inventory and coverage pass.
- seed regenerated; `x bootstrap-cfree` fixpoint holds.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs/language/noise/_section.md

@@ -0,0 +1,13 @@
+---
+id: noise
+title: Noise
+order: 6
+---
+
+Procedural noise — the primitive that terrain, caves, biomes, textures, clouds, wind, and object placement are built on. Every generator is implemented in Q16.16 **fixed point** over an integer permutation hash seeded from an explicit seed, so a given seed reproduces the *exact* same field on every platform and every run: native, headless, and (later) wasm all agree bit-for-bit. That is a real edge over float-based engines, whose worlds can drift subtly across CPUs and break shared-seed multiplayer or replays.
+
+Coordinates are <a href="type-fixed"><code>fixed</code></a> values. The integer part of a coordinate selects a lattice cell and the fraction interpolates within it, so you scale feature size by sampling at a fractional *frequency* (e.g. multiply coordinates by <code>1/64</code>). Outputs are <code>fixed</code> normalised to <code>[-1, 1]</code>; <a href="noise-unit"><code>Noise.unit</code></a> remaps that to <code>[0, 1]</code> when you want a height or a probability.
+
+Pick a generator by feel: <a href="noise-value2"><code>value2</code></a> is cheap and blocky; <a href="noise-perlin2"><code>perlin2</code></a> is the classic gradient noise; <a href="noise-simplex2"><code>simplex2</code></a> is the organic default with fewer directional artifacts; <a href="noise-fbm2"><code>fbm2</code></a> stacks octaves of simplex for natural, detailed fields; and <a href="noise-cellular2"><code>cellular2</code></a> (Worley) gives Voronoi-cell structure for stone, cracks, and biome boundaries. Every sampler is a pure function of <code>(x, y, seed)</code> — no global state, no allocation — so it is safe to call across a whole worldgen pass or a per-pixel fill.
+
+Seed worldgen from its own seed (or a dedicated <a href="ns-Random"><code>Random</code></a> stream), kept separate from gameplay RNG, so generating the world never desyncs the simulation.