The release choice and the pass judgement are internal functions (SelectRelease, JudgePass) and Solve reports each
decision, in order and with the values it compared, to an optional internal ISolveTrace. The tests assert those decisions
with exact equality against the traces the base commit dbcb2a8e produced on the same problems, and call the two functions
directly on states a solve may never reach. Solve's own n < 0 check is deleted (Validate throws the same exception), and the
NonFiniteTarget member doc names the n = 0 case.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A rescaled problem takes five passes on its own merits; the assertion is that the solve does not spin to the cap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
W5 with its columns rotated so every column is the released one; the noise gate across column scales; two small problems
whose release is confirmed by a second pass. Untested at this commit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds BoundedLeastSquaresInput and BoundedLeastSquares.Validate (tolerance, bounds, matrix, target; first failure).
Solve calls it and its inline tolerance, bounds and A^T b finiteness checks are deleted. Untested at this commit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
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>
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>
Batches of small boxed problems whose passes end with the KKT test failing and
the cost unmoved (dependent free columns, releases the factor drops, weak-direction
noise-floor problems). They must report Converged, and land within the documented
gap of an exact active-set enumeration. Pins the stall comparison and the cost
scale in Objective, which no test reached before.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
With b along the smallest left singular vector, |A^T b| is tiny, so the KKT
tolerance (1e-6 of it) sits below the gradient's rounding noise at cond(A) >=
1e6 and can never be met. The solver must stop on the progress rule, in at most
six passes, never at the cap. This is the path the progress rule exists for;
the earlier fixtures reached it only by accident.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Two tolerance behaviours had no observer: the scale taken at the box point
nearest the origin (a 6e-5 release on a 1000-scale problem, in either column
position), and the strict comparisons when the scale and the gradient are both
exactly zero (b = 0 at the origin takes no iteration).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The full-step clamp repairs an out-of-box start on a free column, so the
warm-start clamp was invisible to the old test (Stryker: five survivors). The
new table runs one column with b beyond the box and a cap of 0, where the
solver returns the clamped start untouched, across finite, infinite and NaN
starts and finite and infinite bounds.
Workspace junk (including NaN doubles) is now checked on random boxed problems
that release bounds and use the tolerance, with bit-identical x and equal
iteration counts.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Stryker left the refinement step (removed, negated, or subtracted) unobserved,
and it is unobservable for a reason: refinement with a same-precision residual
does not improve forward error, and Cholesky is backward stable, so the
residual the KKT test reads is already at rounding level. Measured on n=24
problems at cond(A) 1e3..1e6, the solver with and without the step gives the
same iteration counts and the same x error against a Gram-Schmidt QR
reference (both limited by cond(A^T A) * eps).
The refinement pass, its dp buffer and the scratch vector are deleted;
WorkspaceLength returns to 4n^2 + 15n. The doc states where the normal
equations stop being optimal (cost gap below 1e-6 |b|^2 through cond(A) 3e5,
about 2e-5 at 1e6, independent of the scale of A) and that a ridge row beyond
that is the caller's policy.
status is now initialised: the unassigned local made Stryker discard every
mutant inside Solve as a compile error, so 255 mutants were never measured.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- n=24 fully free problems at cond 1e2/1e3/1e4, and A x 1e6 with unscaled
bounds, converge in at most two iterations (a regression to cap-spinning
fails on the count, not on a clock).
- b = 0 from a far start reaches the optimum instead of the start.
- Warm-start clamp table, free duplicate column gets no step, parallel columns
at absolute scales 1e3 and 1e-3, near-singular free sets confirm once and
stop, near-collinear optimum at A x 1e6, and zero allocation on the invalid,
converged and cap paths (first and second call).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
