islaApocalypse-v2/Tools/Scripts/ShapingOracle.cs
beezm 35b4818e9e Phase 2a: the faithful redistribution curve, re-measured against v2's own output
Ports the reference's v5 height curve and shelf-detail passes onto Phase 1's shape
and re-calibrates them against this repo's actual pass-1 distribution. This is the
BASELINE the reshape gets judged against, not the reshape.

Core (engine-free, D-060):
- WorldScale — THE vertical yardstick. One metres/raw number (251), replacing the
  prototype's three duplicate M_PER_UNIT constants and ~20 bare literals. The
  chunk-height coupling it had there is recorded as a DEFERRED vault decision, not
  inherited. RawFromMetres divides, matching the reference bit-for-bit.
- HeightCurve — the 7 bands, the frozen corner-fix blends, the per-seed spike
  normalization, the 24-corner monotonicity sweep that throws and refuses.
  Identity at and below sea, which everything downstream rests on.
- CurveKnots / CurveAnchors — input knots (measured percentiles) and output anchors
  (storm ladder) split apart and both made parameters, so the anchors are A/B-able
  without editing source. The reference's shipped knots are kept beside the measured
  ones as the fidelity yardstick.
- TerrainDetailPass — micro-relief skin plus the shelf-edge KNOT warp (which slides
  K3/K4/K5, not height — that is what keeps monotonicity structural). The crater
  exclusion is ported and inert until the carve lands.

Tools:
- Shaping — pass 2a, producing the two height fields. classify is bit-for-bit the
  raw pass-1 field; render is curved and detailed. Aliased when the curve is off,
  as the reference did. Pass1Result is left immutable so the oracle can compare.
- LandHistogram — the calibration engine AND the diagnostic. The reference shipped
  six knot literals and threw the measuring instrument away; this rebuilds it.
- ShapingOracle + CurveBaselineTool — four automatic checks before anything is
  looked at, and the batch that runs them.

Measured, not assumed:
- Knots re-measured over a 6-seed / 12.8M-sample pool. They differ from the
  reference's by at most 5.6 m of world height, against a 44.7 m per-seed spread —
  the pass-1 port is faithful.
- Oracle all pass, including pass 1 bit-identical to Phase 1's own .f32 dump.
- Band shares land on 60/13/10/5/8/3/1 to 0.00 pp.
- Knots hold across map size: the 8K delta (5.8 m) sits inside seed noise.

The finding the histograms deliver: 83% of land ends below 100 m and 96% below
220 m, with the median column at 13 m. That is the share targets doing exactly what
they say, not a bug — and it is the developer's call, which is why nothing here
reshapes it and the palette was deliberately left mis-fitted rather than recalibrated
to disguise it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DCWNaDZPfTiAy3meGNGgqt
2026-08-20 01:38:10 -04:00

238 lines
9.4 KiB
C#
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

using System;
using System.Collections.Generic;
using IslaApocalypse.Core;
namespace IslaApocalypse.Tools
{
/// <summary>
/// ⭐⭐ THE ORACLE — the automatic correctness checks that let the developer's eye judge ONLY
/// relief.
///
/// ═══ WHY THIS EXISTS AT ALL ═══
///
/// The prototype's single most valuable terrain lesson was not about terrain:
///
/// > *"Five rounds of taste-iteration were safe BECAUSE correctness was not being judged by eye.
/// > Where a future phase has a subjective gate, ask first what the automatic invariant is."*
/// > — `Design - Tooling - Iteration and Batching.md`, "build the oracle before the
/// > taste-iteration, not after".
///
/// The curve is a subjective gate. So before a single render is looked at, four things are
/// proven mechanically:
///
/// (a) REGRESSION curve OFF is bit-identical to Phase 1's pass-1 output.
/// ⇒ the port disturbed nothing upstream.
/// (b) CLASSIFY FIDELITY the classify field is bit-identical to the raw pre-curve field,
/// curve on or off. ⇒ the oracle field is actually an oracle.
/// (c) MONOTONICITY the effective per-seed curve is strictly increasing everywhere.
/// ⇒ no peak has become a pit.
/// (d) BAND SHARES realized land shares match the 60/13/10/5/8/3/1 targets.
/// ⇒ the calibration did what it claimed.
///
/// ⚠ BIT-IDENTICAL MEANS BIT-IDENTICAL. These compare IEEE-754 bit patterns, not values within
/// an epsilon. "Close enough" is how a drift becomes a fact — and the whole point of the `.f32`
/// dump is that two generators agree or their dumps differ.
///
/// ⚠ Any failure fails the TASK, loudly. Nothing here papers over a mismatch: a check that
/// reports "mostly passed" is a check that has stopped working.
/// </summary>
public static class ShapingOracle
{
/// <summary>One check's verdict. <see cref="Detail"/> carries the evidence either way.</summary>
public sealed class Check
{
public string Id; // "a", "b", "c", "d"
public string Name;
public bool Passed;
public string Detail;
public override string ToString() => $"[{(Passed ? "PASS" : "FAIL")}] ({Id}) {Name} — {Detail}";
}
/// <summary>
/// Compare two float fields for BIT equality. Returns the number of differing cells and the
/// first difference found, so a failure is actionable rather than just red.
/// </summary>
public static (long differing, string firstDiff) CompareBitwise(float[,] a, float[,] b, int mapSize)
{
long differing = 0;
string first = null;
for (int x = 0; x < mapSize; x++)
{
for (int y = 0; y < mapSize; y++)
{
int ba = BitConverter.SingleToInt32Bits(a[x, y]);
int bb = BitConverter.SingleToInt32Bits(b[x, y]);
if (ba == bb) continue;
differing++;
first ??= $"first at [{x},{y}]: {a[x, y]:G9} (0x{ba:X8}) vs {b[x, y]:G9} (0x{bb:X8})";
}
}
return (differing, first);
}
/// <summary>
/// (a) REGRESSION — with the curve off, shaping must return the pass-1 field untouched.
///
/// ⚠ This is the WEAKER, always-available half of check (a): it proves pass 2 is a no-op when
/// gated off. The stronger half — that pass 1 ITSELF still matches Phase 1 byte-for-byte — is
/// <see cref="RegressionAgainstDump"/>, which needs a Phase-1 `.f32` on disk.
/// </summary>
public static Check RegressionCurveOff(Pass1Result p1, Pass2Result offResult)
{
var c = new Check { Id = "a", Name = "regression: curve OFF == pass-1 output" };
if (offResult.CurveOn)
{
c.Passed = false;
c.Detail = "the result handed in was generated with the curve ON — wrong variant.";
return c;
}
var (differing, firstDiff) = CompareBitwise(p1.Height, offResult.Height, p1.MapSize);
bool aliased = offResult.FieldsAreAliased;
c.Passed = differing == 0;
c.Detail = c.Passed
? $"bit-identical over {(long)p1.MapSize * p1.MapSize:N0} cells" +
(aliased ? " (and the fields alias one array, as the reference did)" : "")
: $"{differing:N0} cells differ — {firstDiff}";
return c;
}
/// <summary>
/// (a) REGRESSION against a Phase-1 `.f32` dump — the cross-run half.
///
/// ⚠ A MISSING DUMP IS NOT A PASS. It is reported as INCONCLUSIVE and the caller says so; a
/// check that silently succeeds when its input is absent is worse than no check, because it
/// buys confidence that was never earned.
/// </summary>
public static Check RegressionAgainstDump(float[,] current, float[,] phase1Dump, int mapSize, string dumpPath)
{
var c = new Check { Id = "a", Name = "regression: pass-1 == Phase-1 .f32 dump" };
if (phase1Dump == null)
{
c.Passed = false;
c.Detail = $"INCONCLUSIVE — no readable Phase-1 dump at {dumpPath} for this seed/size. " +
"Not counted as a pass; set ISLA_PHASE1_SOURCE to a batch that has one.";
return c;
}
var (differing, firstDiff) = CompareBitwise(phase1Dump, current, mapSize);
c.Passed = differing == 0;
c.Detail = c.Passed
? $"bit-identical to {dumpPath} over {(long)mapSize * mapSize:N0} cells"
: $"{differing:N0} cells differ from {dumpPath} — {firstDiff}";
return c;
}
/// <summary>
/// (b) CLASSIFY FIDELITY — the classify field is the raw pre-curve field, bit-for-bit, with
/// the curve on or off.
///
/// This is the invariant every later phase's oracle rests on: biomes and water will classify
/// from this field, so if it has drifted by even one ulp the "md5-identical across shaping
/// changes" guarantee is gone before it is ever used.
/// </summary>
public static Check ClassifyFidelity(Pass1Result p1, Pass2Result p2)
{
var c = new Check { Id = "b", Name = "classify field == raw pass-1 field" };
var (differing, firstDiff) = CompareBitwise(p1.Height, p2.HeightClassify, p1.MapSize);
c.Passed = differing == 0;
c.Detail = c.Passed
? $"bit-identical over {(long)p1.MapSize * p1.MapSize:N0} cells (curve {(p2.CurveOn ? "ON" : "OFF")})"
: $"{differing:N0} cells differ — {firstDiff}";
return c;
}
/// <summary>
/// (c) MONOTONICITY — recorded rather than re-run.
///
/// <c>HeightCurve.AssertMonotonic</c> THROWS on violation and is called inside
/// <see cref="Shaping.Shape"/>, so reaching this code at all means the sweep passed. The
/// check exists so the oracle table states it explicitly instead of leaving the strongest
/// guarantee implicit in the absence of a crash.
/// </summary>
public static Check Monotonicity(Pass2Result p2)
{
var c = new Check { Id = "c", Name = "curve strictly monotonic (24-corner sweep)" };
if (!p2.CurveOn)
{
c.Passed = true;
c.Detail = "curve off — nothing to prove (identity is trivially monotonic).";
return c;
}
string note = p2.Notes.Find(n => n.Contains("Monotonicity assertion passed"));
c.Passed = note != null;
c.Detail = note ?? "no monotonicity confirmation recorded — AssertMonotonic did not run.";
return c;
}
/// <summary>
/// (d) BAND SHARES — the realized land shares against the M3 targets.
///
/// ⚠ MEASURED ON THE RAW (INPUT) DISTRIBUTION, because that is where the knots cut. The
/// shares are exact by construction IF the quantile machinery is right — so this check is
/// really a proof that <see cref="LandHistogram"/> measured what it claimed, which is the one
/// thing the reference could never verify about its own knots.
/// </summary>
/// <param name="tolerancePercentagePoints">
/// Allowed absolute deviation per band, in percentage points. The knots come from the SAME
/// histogram, so agreement is limited only by in-bin interpolation — tenths of a point, not
/// whole ones.
/// </param>
public static Check BandShares(LandHistogram raw, CurveKnots k, double tolerancePercentagePoints)
{
var c = new Check { Id = "d", Name = "realized land band shares == 60/13/10/5/8/3/1 targets" };
double[] realized = RealizedShares(raw, k);
double worst = 0.0;
int worstBand = -1;
for (int i = 0; i < realized.Length; i++)
{
double d = Math.Abs(realized[i] - CurveKnots.BandShareTargets[i]);
if (d > worst) { worst = d; worstBand = i; }
}
c.Passed = worst <= tolerancePercentagePoints;
c.Detail = $"worst band '{CurveKnots.BandNames[worstBand]}' off by {worst:F3} pp " +
$"(tolerance {tolerancePercentagePoints:F2} pp); realized " +
string.Join("/", Array.ConvertAll(realized, v => v.ToString("F2")));
return c;
}
/// <summary>
/// The fraction of land, in percent, falling in each of the curve's seven INPUT bands.
/// Band edges are sea, K1..K6, +∞.
/// </summary>
public static double[] RealizedShares(LandHistogram raw, CurveKnots k)
{
float[] edges = { raw.SeaLevel, k.K1, k.K2, k.K3, k.K4, k.K5, k.K6 };
var shares = new double[7];
for (int i = 0; i < 6; i++)
shares[i] = raw.FractionBetween(edges[i], edges[i + 1]) * 100.0;
shares[6] = Math.Max(0.0, (1.0 - raw.FractionBelow(k.K6)) * 100.0);
return shares;
}
/// <summary>Render the whole oracle as a markdown table for the INDEX and the report.</summary>
public static string ToMarkdownTable(IEnumerable<Check> checks)
{
var sb = new System.Text.StringBuilder();
sb.AppendLine("| | Check | Result | Evidence |");
sb.AppendLine("|---|---|---|---|");
foreach (Check c in checks)
sb.AppendLine($"| `{c.Id}` | {c.Name} | **{(c.Passed ? "PASS" : "FAIL")}** | {c.Detail} |");
return sb.ToString();
}
}
}