A bound is released only when its violation exceeds both kktTol and the rounding
bound of its double gradient (Higham gamma_(n+1)); the cost-progress rule now judges
only passes that release nothing, and a release resets fMark. Fixes a cold start
stopping Converged with a violated bound (Soul W5, finding
aetheria-release:finding:cut-thrust-allocator-core.s1.solver-converged-early).
Tests: the W5 problem (cold == re-solve at 1e-12 and 1e-9) and the seed-141 noise
case. Builds; verified on Yggdrasil in the following report.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Imagination pass (session self-2026-10-10-ag): the cost-progress stop at
BoundedLeastSquares.cs:275 reports Converged with a bound still violated;
fix follows Lawson-Hanson. Quaternion algebra surface and the
mul(float4,float4) overload trap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Eyes pass, 2026-10-10: inverse CSG and shape-program synthesis, primitive
fitting front ends, differentiable SDF/CSG fitting, measured LLM spatial
reasoning with a target, hard-surface detail, and the VibeGeometry record.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cut cultmath-release (campaign cultlib-gaps). Soul passes 1-4; the last docs edit is the one pass 4 prescribed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Stale logs are deleted before launch; the results file must be newer than the launch. All commits build (no source change).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
packages/cultmath/scripts/unity-smoke/ holds the good compute shader, the
deliberately broken control, the EditMode test (FindKernel, dispatch, readback),
and run-unity-smoke.ps1, which builds a scratch project outside the repo with a
file: path to the package, takes the Unity path as a parameter, requires all
three tests to pass, and deletes the project. The package README's Unity section
names it as the pre-tag step. Not -nographics: GetComputeShaderMessageCount
reports 0 for a broken shader until FindKernel or Dispatch compiles it. Ran once
on Starfire with Unity 6000.3.24f1: 3 passed, control rejected. Builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Asserts per output (value, and each snoise_grad component) that it differs
from 0.3.0 on the tie set, and pins the 0.4.0 values as goldens beside the
0.3.0 ones, so reverting either math.cs:614 or :720 alone fails the test. The
comment says inputs within a few float32 ulps of the set can tie too. Not yet
run when committed; the next commit-free verification run follows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CHANGELOG and package README: the noise enclosures rest on SNOISE_LIPSCHITZ,
SNOISE2_LIPSCHITZ and SNOISE_HESSIAN, all empirical with a margin; the README
bullet no longer says "with a proof" or "every point". Breaking: inputs within a
few float32 ulps of the tie set can also change (probe: none at 1e-5 or more,
a handful at 1e-6 near magnitude 30). Text only; builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CHANGELOG and README no longer claim iv_snoise_ball and iv_fbm_ball enclose
every input unconditionally: they rest on SNOISE_LIPSCHITZ, an empirical bound
with a margin, and the tests sample. Docs only; builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Breaking entry now names the whole tie set (inputs whose y-x and z-x are
whole numbers, every integer point included, change by up to 0.776), with
examples and a consumer warning; design.md and the NoiseGradTests comment drop
the narrower origin-and-(t,t,t) wording. Docs and test comment only; builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Inputs whose coordinate differences y-x and z-x are whole numbers move from the
0.3.0 value; inputs off that set keep its bits. The goldens were computed by
the cultmath-unity-v0.3.0 tag's tracked CultMath.dll. Builds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copied from artifacts/cultmath-unity-build of the release script's first run
(netstandard2.1, non-incremental, Source Link off, ContinuousIntegrationBuild).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bumps CultMath.csproj, the Unity package.json, the org.gamecult.cultlib
dependency and the semver-policy paragraph to 0.4.0. CHANGELOG states the snoise
tie rule under Breaking and the additions since 0.3.0. The tracked DLL and pdb
still read 0.3.0 at this commit, so the release script's parity check fails here.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The release flips at 1.05x and 0.95x of the computed relative gradient at four
weights either side of 1e-6; all-free conditioned solves change between the
default and 1e-12; the omitted tolerance releases at 1.1e-6 and not at 0.9e-6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Default and explicit-default are bit-identical; a light term under a heavy row
and an allocator-shaped problem are start-dependent at the default and solved
tight at 1e-12; invalid tolerances leave x untouched; the noise-floor fixtures
stop on progress at 1e-12 and 1e-30 (at most 6 iterations); the tolerance path
allocates nothing; the original ten-parameter signature still exists.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
New overload with kktRelativeTolerance before maxIterations; the existing
signature forwards with the new public DefaultKktRelativeTolerance (1e-6), which
replaces the private constant. A tolerance that is not finite and positive is
InvalidInput with x untouched. The default path is bit-identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Records status, iteration count and every float bit of x for the A54/B5 fixture,
three seeded random problems and a Duel allocator problem, taken at f7ab6212.
The kktRelativeTolerance cut must keep these literals unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Reduced affine arithmetic (centre, one shared symbol, one error term) for CultMath:
af_* ops, af_snoise / af_fbm with gradient-plus-curvature forms, af_frustum_*, the
2D iv_snoise_ball, and the centred warp bound. Mirrored bit for bit in
CultMath.Affine.hlsl (and the Unity copy), lowered into CultMath.glsl, with fixture
families that check enclosure on device, including small-reach and C#-centre legs.
Scenario (c) is a deep well an order of magnitude past the fog band; affine beats the
centred interval there (3.57x / 3.46x vs 3.20x / 3.12x), asserted, not skipped.
Soul held cut-affine-forms (s1-s3) and cut-affine-shaders (s1-s3).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Re-measured at c01ff025 with IntervalSkipHalvesEvaluations: (c) affine 3.57x and 3.46x, c-band 3.95x
and 3.54x, efficiencies and range-ceiling ratios as printed; the (b) rows and the DeepWellCullMap
figures (4.23, 3.60, 2.02) are unchanged. The sentence that justified deleting the affine detail
ball is gone with the deletion.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Deleting the affine gate's detail ball (51a4034c) gave up 6 of 5933 provable culls inside the fog
(first: (b) inside the fog [affine]); the gate never gives up a provable cull, and a 1.4% probe
saving does not buy breaking that. The detail ball is back in the affine gate, and the gate test
now includes the fog interior, the only field where the interval gate's detail ball decides, so
saturating either gate's detail ball fails the test.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Cells per ray 4.23 interval, 3.60 affine, 2.02 oracle, the most 44, mean transmittance 0.278,
measured at the affine gate fix; the interval figure (4.51, 45) was already stale: the interval
path is unchanged by that fix.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Two families are appended (256 cases each, reach |axis| + radius of 1e-6 to 1e-3, the centre in
[-50, 50] or [-1, 1]); the earlier families are byte-identical. In the main families the
second-order remainder is at least 0.057 of e even without the allowance; here the float32
allowance is most of e, so a device whose gradient error exceeds it fails the enclosure check.
Their check text adds the centre leg for every af_snoise and af_fbm family: the fixture's own
x0 (C#'s value at the centre) must lie in the device form's [x0 - e, x0 + e], so a device whose
noise is wrong everywhere no longer passes against its own reference.
AssertSampledForm counts the cases that need the allowance (32 of 256 af_snoise, 3 of 256
af_fbm on C#'s own results) and the small families fail unless there is at least one.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
EveryMirrorFunctionMatchesCSharpMath draws centre, axis and radius from [-100, 100], far past the
0.15 or so below which the centred form wins, so an HLSL edit of the choice survived it.
AffineSnoiseMirrorMatchesCSharpBitForBitOnBothSidesOfItsBranch draws 20000 slices (axis and
radius log-uniform 1e-3 to 0.5), asserts both branches are met more than 2000 times each, and
compares af_snoise (and af_fbm every eighth) bit for bit.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
dxc marks compares fast, so `lo == lo ? lo : -inf` was deleted under default flags and an
overflowed form ranged to NaN on HLSL targets. The guard now tests (asuint(lo) & 0x7fffffff)
> 0x7f800000, in HLSL, the Unity copy, math.Affine.cs and the regenerated CultMath.glsl.
dxc 1.9.0.5402 -T ps_6_0 default flags: before, no fcmp/select/and in the af_range body;
after, and/icmp ugt/select to +-inf for both endpoints (also under -Gis).
AffineRangeMirrorMatchesCSharpBitForBitAtTheFloatExtremes pins the mirror at zero to the
float maximum, infinities and NaN components.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The affine gate's best-case detail ball decided no probe on the gate test's slices and one
of the deep well's 8074 oracle-empty cells (11 of 11096 in c-band); skipping it saves the
probes: (c) affine 3.57x -> 3.62x and 3.46x -> 3.54x, c-band 3.95x -> 3.96x and 3.54x -> 3.60x.
The interval gate keeps its detail ball. Saturating it there (measured, then reverted)
takes row (b), inside the fog, from a cull fraction of 0.663 to 0.585 and a cost ratio
of 1.16x to 0.98x, so the term is live; the finding's premise holds for the affine gate only.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Slice builds each phase's coverage argument and the detail argument (centre,
axis, radius in noise space); AffineBound and its gate read them whole.
AffineWarpStaysEnclosed checks the rays against those arguments and their
derivation from CentredWarp; AffineCompositionIsTighterThanItsRanges sees the
shared-eps composition and the intersection.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The 2D ball carries no evaluation allowance, like iv_snoise_ball(float3); the
caller that rounds the flow's point adds 2^-14 + L2 2^-20 (|centre|_1 + radius),
the convention af_snoise carries for the 3D point.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Upstream Ashima's step compares are cyclic, so where x0's components compare equal
(the origin, every (t,t,t), all lattice diagonals) the corners stop being a simplex
and snoise jumps (about 0.77 near the origin, up to 1.56 further out). The z-x compare
is now strict, giving Gustavson's simplexnoise1234 order in C#, both HLSL copies and
the generated GLSL. Off the z=x tie set values are bit-identical. SNOISE_LIPSCHITZ
re-pinned 10.099261 -> 7.9640074. A value change: next cultmath-unity release is
Breaking (follow-up snoise-tie-release). Soul s1 held (cut-snoise-tie.h2).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Soul s1 (cut-snoise-tie) measured 1.55 to 1.56 at magnitudes 64 and above;
0.77 holds only near the origin. Comment and design text only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
