islaApocalypse-v2/Core/Scripts/GenerationScale.cs

using System;

namespace IslaApocalypse.Core
{
	/// <summary>
	/// ⭐ THE SCALING DISCIPLINE, IN CODE. → `Design - Tooling - Scaling Discipline.md`, D-002.
	///
	/// ═══ THE RULE ═══
	///
	///     NOTHING IN THE GENERATOR USES A RAW PIXEL NUMBER.
	///     Every distance, radius, threshold and noise frequency is a FRACTION OF MapSize,
	///     or derived from ScaleFactor.
	///
	/// ═══ WHY IT IS A RULE AND NOT A PREFERENCE ═══
	///
	/// These failures are SILENT. A hardcoded pixel count does not throw — the generator still
	/// runs and still produces a map, and the map is just subtly wrong in ways that are hard to
	/// attribute. The prototype learned this the expensive way: moving to 1:1 scale made the world
	/// four times wider in pixels while the noise settings stayed put, and the terrain became
	/// "TV static" — the same features packed into a quarter of the space. The same transition
	/// broke a long tail of checks like "look 10 pixels away to see if there's water", which
	/// quietly started measuring a quarter of the distance they used to.
	///
	/// In the developer's own words: "These kind of numbers I'm getting tired of changing because
	/// of map size. We need to make sure these scale!"
	///
	/// ═══ ⚠ ONE BASELINE, DELIBERATELY ═══
	///
	/// The prototype ran TWO reference scales side by side: /1024 for noise and map drawing,
	/// /4096 for town counts and road repulsion. Neither was wrong, but a reader — or a resize —
	/// had to know which applied where, and the design doc's standing instruction is to "pick
	/// deliberately and say which". This repo picks: ONE baseline, 1024, for everything. If a
	/// second is ever genuinely needed it gets its own named type and its own reason, not a
	/// bare divisor at a call site.
	///
	/// ═══ ⚠⚠ THE TIGHTENING OVER THE REFERENCE — NORMALIZED ADDITIVE OFFSETS ═══
	///
	/// The reference decorrelated noise layers with offsets in RAW PIXELS — GetNoise2D(x + 1000, …),
	/// GetNoise2D(x * 1.5f + 5000, …). Because frequency already carries a 1/MapSize normalization,
	/// a pixel-space offset does NOT hold still across map sizes: 1000 px is 1.0 normalized units
	/// at 4K but 0.5 at 8K, so a resize silently samples a DIFFERENT SLICE of the noise field. The
	/// feature scale is preserved; the REALIZATION is not. The same seed at two map sizes gives two
	/// different worlds, for no reason anyone wrote down. (Found by the chat1/00 reference
	/// inventory, §4.2 — where it also corrected the vault's standing diagnosis that the detail
	/// noises were unscaled. They are scaled; it is the OFFSETS that are not.)
	///
	/// So in this repo: DECORRELATION OFFSETS ARE DECLARED IN MAP WIDTHS, via
	/// <see cref="OffsetInMapWidths"/>. Never write a bare pixel constant into a noise coordinate.
	///
	/// No noise is built this phase. The convention is established before the generator exists so
	/// that nothing is ever added unscaled — which is the only time this rule is cheap to hold.
	/// </summary>
	public readonly struct GenerationScale
	{
		/// <summary>
		/// The reference map width the scaling baseline is anchored to. The world's noise looked
		/// right at 1024 columns; every larger map stretches its features back out to match.
		/// </summary>
		public const int BaselineMapSize = 1024;

		/// <summary>
		/// The map's side in columns. 1 column = 1 metre (D-002, 1:1 scale).
		///
		/// ⚠ A GENERATION PARAMETER. Never a constant, never baked in. The rewrite's default target
		/// is 10k x 10k, configurable 8k–12k (D-056) — but that is a DEFAULT chosen by config, and
		/// no code below this line may assume it.
		/// </summary>
		public readonly int MapSize;

		public GenerationScale(int mapSize)
		{
			if (mapSize < BaselineMapSize)
				throw new ArgumentOutOfRangeException(nameof(mapSize), mapSize,
					$"MapSize must be at least the {BaselineMapSize} baseline.");
			MapSize = mapSize;
		}

		/// <summary>
		/// MapSize / 1024. Divide a baseline-tuned noise frequency by this to hold feature size
		/// constant in METRES as the map grows. A mountain range is the same physical size at any
		/// map profile — which is what makes the map-size setting safe to change at all.
		/// </summary>
		public float ScaleFactor => MapSize / (float)BaselineMapSize;

		/// <summary>
		/// A distance, from a fraction of the map. Use this instead of writing a pixel count.
		/// <c>Fraction(0.75f)</c> is "three quarters of the way across", at every map size.
		/// </summary>
		public float Fraction(float fractionOfMap) => fractionOfMap * MapSize;

		/// <summary>
		/// The inverse: what fraction of the map a distance in columns represents. For turning a
		/// measured pixel distance back into a scale-free constant before you commit it to code.
		/// </summary>
		public float AsFraction(float distanceInColumns) => distanceInColumns / MapSize;

		/// <summary>
		/// A noise frequency, from one tuned at the 1024 baseline.
		/// <c>NoiseFrequency(0.004f)</c> reproduces the reference's base terrain frequency at any
		/// map size. → the reference's <c>0.004f / scaleFactor</c>, chat1/00 §2.1.
		/// </summary>
		public float NoiseFrequency(float baselineFrequency) => baselineFrequency / ScaleFactor;

		/// <summary>
		/// ⭐ A decorrelation offset for a noise coordinate, declared in MAP WIDTHS.
		///
		/// <c>OffsetInMapWidths(0.25f)</c> shifts the sample by a quarter of the map at EVERY map
		/// size — so two noise layers stay exactly as decorrelated at 12K as they were at 8K, and
		/// the same seed produces the same world shape at any size.
		///
		/// ⚠ This is the one place a decorrelation offset may come from. A bare pixel constant in a
		/// noise coordinate is the bug described in this type's summary; there is no correct value
		/// for one.
		/// </summary>
		public float OffsetInMapWidths(float mapWidths) => mapWidths * MapSize;

		public override string ToString() => $"GenerationScale(MapSize={MapSize}, ScaleFactor={ScaleFactor:F3})";
	}
}