v0.5.0 - eight silent-value fixes, a kink-aware Gauss-Legendre rule, demos 4 and 5

Eight inputs returned a wrong number without an error; each now returns the right value or raises. They were found by merging the open branches, reviewing the result, and building two demos in configurations where the wrong and the right value differ.

Breaking

• System.expand raises ValueError for diag_R=True when R has off-diagonal entries (a dense ExplicitR(iso_R=False)) and for diag_C=True when C has them (a GeneralKappa2 mixing matrix, a matrix ConstantImpulse / CustomImpulse, or a dense R). System.propagators refuses diag_C=True with an off-diagonal C, and a t_max above t_max_cache. Pass diag_R=False / diag_C=False; in this release an off-diagonal C still needs c_closed_form_only=True and a closed form.
• The L0 evaluators raise ValueError when a term's R indices contradict the propagator cache's R (two-index terms against a scalar R, index-free terms against a matrix R, an absorbed R with two indices).

Values that change

defect	affected inputs	size before the fix
callable non-local coupling evaluated at one leg order	callable κ not symmetric in its leg points	53 % to 130 % (the tests' generic kernel)
diag-C fast path ignored pinned observable labels	qmc_scalar, qmc, nquad under diag_C	17 % to 78 % at order 2
off-diagonal R and C dropped under the L1 defaults	dense ExplicitR, mixing GeneralKappa2, matrix white noise	order-0 ⟨φ₀φ₁⟩ = 0 instead of 0.1178
DiagonalA with a callable rate	components that differ away from t = 0 and 1; R below t = 0	23 % on R₁₁; 13 to 1400 times too large below 0
System.t_min not passed to the integrators	t_min ≠ 0 through Expansion.evaluate, sweep, the CLI	×1.27 on an order-1 three-point function
iso_C=True multiplied every C by N	iso_C=True, N ≥ 2	×2 at orders 0-1, ×4 at order 2 (N = 2)
a repeated R pair reused the first R's indices	L0, matrix R, a vertex with two ψ legs	totals off by 2.1 to 5.8 times, one sign flip
L0 R indices against the cache's R	L0 flags that contradict the cache	×3.7 to ×27

One crash is fixed: a callable coupling with a matrix R on a zero-dimensional diagram failed on qmc_vectorized, gauss_legendre and nquad.

Demos 1-3 reproduce their stored numbers.

Gauss-Legendre on white noise

gauss_legendre splits the time domain where the integrand is kinked (the white-noise C diagonal; two parents of a vertex with several ψ legs), recursively. Demo 4 against its exact hierarchy:

channel	before, 8 / 16 / 64 nodes	now, 8 / 16 nodes
white pulses, FF	4.5e-3 / 1.2e-3 / 7.9e-5	2.0e-13 / 7.0e-16
exponential pulses, FF	1.2e-5 / 8.2e-7 / 3.5e-9	6.9e-11 / 1.1e-15

Demos 4 and 5

• Demo 4 (examples/demo4): two-component compound-Poisson noise whose cumulants are symmetric only under joint permutations of (component, point) pairs, and an F with no index symmetry. Level A against Campbell's theorem, level B against an exact moment hierarchy. On the code before the leg-order fix it is 46.8 % off.
• Demo 5 (examples/demo5): white noise with a matrix amplitude, t_min = 0.5, orders 0-4 on every integrator (Gauss-Legendre to 1.4e-15), and multiplicative noise at L0 with scalar and matrix R (to 6.2e-16). On the pre-fix commits it fails by factors 2 to 8 (iso_C) and by 64 % (diag-C fast path).
• examples/reference/ito_moments.py: the exact moment hierarchy of a polynomial Itô SDE with jumps, used as the reference; it imports nothing from sft-wick.

Every fix, with its measurements and the tests that lock it, is in CHANGELOG.md.