CellularAndSminGradTests.cs changes for cut 2a-i's pass-2 findings:
- FeaturePoint and the two other independent pcg3d/pcg4d call sites now hash
int3(cell)/int4(int3(cellId), 0) to match production's integer-hashing fix
(previous commit); the oracle would otherwise diverge from cellular() and
every comparison test would fail.
- New JitterAndIdPassAChiSquareUniformityTest: jitter.x/y/z and id each pass
a chi-square uniformity test (16 bins, 15 df, critical value 37.7 at
alpha=0.001) over 200,000 cells. This is the behavioural pin for the
hashing fix; reverting to pcg3d(cell)'s float-bit hash fails it.
- Two false comments corrected to the measured facts: the
NearestFeatureWins test's claim that |component|=2 offsets are "provably
never the true nearest" is false (their lower bound is 1, not sqrt(3); Soul
measured 3 F1 wins in 2,000,000 samples), and SearchReachesEveryAxis...'s
claim of "lower bound is exactly 2" and "zero hits in 3,000,000" is also
false (lower bound is 1; Soul measured about 1 hit in 28,000).
- PrunedSearchIsBitIdentical... renamed to
...BelowTwoToTheTwentyFifth and its comment states that domain explicitly
(design.md, previous commit).
- GradientTolerance split into SminGradientTolerance (2.5e-4f) and
CellularGradientTolerance (1.3e-3f): the two functions have different
curvature (smin_grad is exactly quadratic in a band; cellular's F1/F2 are
real distance fields), so a shared 3e-4 tolerance carried almost no
headroom over either's actual measured noise (re-measured here at 5.3e-4
for smin's old eps, 2.0e-4/2.6e-4 for cellular), not the "~6e-5" the old
comment claimed. smin_grad's eps moves from 0.001f to 0.01f (no truncation
cost there; dilutes float noise 8x) with a matching widened seam-exclusion
band so p +/- eps cannot step across the seam the check at p already
avoided. cellular's eps stays at 0.0005f (a small sweep found it already
near the truncation/noise sweet spot for a real distance field). Both new
tolerances keep >5x headroom over their own measured worst case and both
still die under a 1.01x-scale mutant on the returned gradient — smin down
to 1.0005x, cellular (M9/M16-style, either gradient) down to 1.002x.
- GradientMatchesCentralDifferencesAwayFromTheF1EqualsF2Seam also excludes a
band around F2 = F3 now, not just F1 = F2: the second-nearest feature's
identity swaps there too, a genuine kink the old test could silently hit
(an unfiltered large sample found a 0.78 edge error this way, swamping the
~2e-4 the filtered test actually sees).
- AxisAlignedRadiusTwoOffsets and FindPointRealizingOffset are extracted from
SearchReachesEveryAxisAlignedRadiusTwoSlice as internal fixtures so
HlslSourceCompatibilityTests (next commit) can reuse the same construction
instead of a second copy.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Stryker (mtp, --since 29e50ad3) found 261 survivors / 24 timeouts across the
whole repo; triaging the ones in this fix batch's own new code (cellular's
loop, its prune, cellular_unit) surfaced one real gap disguised as
collateral timeout noise from a batched infinite-loop mutant: dz < 2 (and
the dy/dx equivalents), silently dropping the +2 layer on one axis alone,
passed every existing random-sampled test. That is not the sweeps being
weak: the exactness proof's own margin (F1, F2 always < sqrt(3), radius-2
cells' lower bound is exactly 2) makes a genuine win from any single
|offset component| = 2 cell astronomically rare over uniformly random
points (measured zero hits in 3,000,000 samples).
SearchReachesEveryAxisAlignedRadiusTwoSlice constructs a real example for
each of the six axis-aligned radius-2 offsets directly: for a candidate
integer cell, build the query point that is the closest point of that
cell's own unit cube to the target neighbour's actual jittered feature.
That construction finds a hit for every offset within a couple thousand
candidate cells, deterministically. This kills dz < 2 / dy < 2 / dx < 2 (all
manually confirmed).
Also manually confirmed via Stryker's own reported mutants: the dz--/dy--/
dx-- (increment-to-decrement) mutants are genuine infinite loops, correctly
caught as timeouts. The f2 >= 0.0f mutant (disabling the F2 = 0 guard) is
killed by the existing NoNaNAtIntegerPointsNear2To24 test. The remaining
survivors/timeouts in the new code (the lower-bound prune's >>= vs >
comparison, its arithmetic terms, and cellular_unit's >> vs >>> shift on an
already-unsigned operand) are equivalent mutants: since the per-axis lower
bound is always 0 or 1 by construction, and pruning is a proven-safe
optimization that never changes cellular's output, no mutation confined to
"when does the prune trigger" can be observed by any value-based test.
Verified by hand: disabling pruning entirely reproduces the unpruned
reference bit for bit, as it must.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Cut 2a-i fix batch 1, findings F1, F3, F4, F5, F6, F7 (Soul falsification of
the first landing):
- cellular now searches a provably exact 5x5x5 neighbourhood (radius 2), not
3x3x3: 3x3x3 missed the true F2 for 9 of 400,000 sampled points. The
comment in math.cs proves 5x5x5 always contains the true F1 and F2 (both
are always < sqrt(3), every excluded cell is always > 2). A cheap
position-only lower-bound prune skips the pcg3d/jitter/distance work for
cells that provably cannot beat the running F2, without changing any
output bit (F6).
- F2 = 0 now gets the same degenerate-gradient guard F1 = 0 already had,
fixing a NaN that was reachable past |p| of about 2^23 (F3).
- id now comes from pcg4d over the winning cell alone, never from the pcg3d
hash that produced that cell's jitter, so id and jitter are no longer
deterministic functions of each other (F4).
- cellular_unit maps the top 24 bits instead of a plain divide-by-2^32, so
the mapping is always inside [0, 1); the old formula rounded
0xFFFFFF80..0xFFFFFFFF up to exactly 1.0f (F7).
- Central-difference tolerances tightened from 0.02 to 3e-4 (measured float
noise at the committed eps is about 6e-5), and cellular's own gradient
test now differentiates cellular's RETURNED .w values instead of a
duplicate FindF1F2 search, so a mutant that corrupts only the returned
gradient can't hide behind an independently-computed value (F1, F5).
- The struct-mirror comparison in HlslSourceCompatibilityTests now checks
field name and declared type in declaration order, not just position, so
a reordered-and-relabelled mirror struct (H2) or a renamed/retyped one
(H3) both fail the walk instead of surviving on values alone (F2).
- FindF1F2 widened to a true independent 7x7x7 brute-force oracle; new tests
pin production's exact agreement with it, the pruned search's bit-for-bit
equivalence to an unpruned 5x5x5 reference, no NaN at 2^24, and id's
correlation with jitter under 0.05.
- Comment/doc fix: smin_grad is Quilez's quadratic-polynomial smin, not
cubic.
Both shaders/CultMath.hlsl and unity/.../Shaders/CultMath.hlsl stay
byte-identical.
Manually verified mutants M9 (grad1 x 1.01), M16 (grad2 x 1.01), and M18 (id
from hash.y) all die against these tests; H2/H3 struct mutants die via the
new comparator tests. Local dotnet test: 191/191 passing. FXC cs_5_0 compile
of a scratch kernel using both cultmath_smin_grad and cultmath_cellular
succeeds.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
dotnet test found two failures where a fixed-decimal-place Assert.Equal
snapped two numbers that already agreed to within about 2e-5 to different
rounded digits, because their true difference straddled a rounding boundary
(0.824987829 vs 0.825006127 -> 0.82 vs 0.83). Both were central-difference
checks (smin_grad's own invariant-8 property test, and cellular's gradient-
matches-central-differences test), not a correctness bug in the primitives.
AssertWithinTolerance replaces the precision-based assertion for both central-
difference comparisons with an explicit absolute-difference check, which
states the actual claim (agreement within the finite-difference step's own
truncation error) without the rounding-boundary artifact. Also dropped an
unused local function (F2At) the build flagged.
Found by the third ygg-verify.sh test run on 77c3580.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Inside namespace CultMath.Tests, unqualified `Random` resolves to the
enclosing CultMath namespace's own Random struct (design.md: "Random is
CultMath's own xorshift32") before `using System;` is even considered, so
`new Random(seed)` in the new test file was constructing CultMath.Random,
which has no NextSingle(). HlslSourceCompatibilityTests.cs already works
around this the same way (`new System.Random(...)`); CellularAndSminGradTests
now qualifies every use the same way.
Found by the second ygg-verify.sh build/test run on e613ed6.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
CellularAndSminGradTests pins the rules the cut map calls out, each against a
mutant it must kill:
- smin's value never exceeds the ordinary minimum, and matches min with that
input's own gradient once |a.w - b.w| exceeds k on either side.
- smin_grad's gradient matches central differences of its own .w at random
seeded points (the formula is exact, not an approximation, so no exclusion
band is needed away from the |a.w-b.w|=k seam itself).
- A direct case falsifies blending the gradient with a constant 0.5 or with
(1 - h) instead of h.
- cellular's F1 never exceeds F2, matched against an independent brute-force
Worley search that shares only the public pcg3d hash, not cellular's control
flow.
- The true nearest feature wins even when an independent search finds it in
the (1, 1, 1) neighbour offset, the last one cellular's own loop order
visits, ruling out a "first found" or early-exit mutant.
- The gradient at a feature point is finite zero, not NaN.
- Both nearest and edge gradients match central differences away from a band
around the F1 = F2 set, where the nearest and second-nearest identity swap
discontinuously.
- id is constant across small moves that stay in the same feature's region and
differs from a distant cell's id.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
