Pin the symbolic quadratic coefficient against regressing (#771)

[Rafael-SOWNet]

Closes #771 — as already fixed, with the coverage it was missing.

#771 reported that integrating a quadratic in x^2 answers NaN as soon as the x^2 coefficient is a symbol, in the rational and the radical family alike. Re-measured against master, none of it reproduces: PR #774 landed after the issue was filed and fixed all seven of its cases.

What is missing is the coverage. #774's own tests do not reach the radical family, the squared-symbol forms, or the shape reached through a substitution — three of the four families #771 names. This adds 18 cases that do.

Verified to have teeth

Not assumed. Run against 4ab9448c, the commit before #774:

	before #774	on master
all cases	16 of 18 fail	18 pass
TheSymbolicAnswerCarriesNoNaN alone	7 of 7 fail	7 pass

The two that do not fail on the old commit are radical cases whose specialised form was already right — which is exactly why the symptom the issue actually reports, the symbolic answer collapsing to NaN under Simplify, is asserted separately. One degenerate branch dividing by a literal zero is enough to take the whole piecewise with it.

Two things the tests do deliberately

Coefficients are substituted after the symbolic integration, never before. Specialising first measures the numeric-coefficient path, which was never broken — and that is precisely what hid this. The answer is a piecewise split on the sign of the symbolic coefficient, so what needs checking is that the branch selected for a given sign is the right one:

var antiderivative = integrand.ToEntity().Integrate("x");   // symbolic first
var specialised = antiderivative.Substitute("b", 2.0).Substitute("a", 3.0).Simplify();
// then differentiate back and sample

Both signs of the symbolic coefficient and both signs of the constant are covered, since they select different branches.

Sampled rather than compared symbolically. These antiderivatives are logarithms and arctangents of surds in the coefficients; asserting a symbolic zero would test the simplifier's reach rather than the integrator's answer. A point where either side leaves the reals or goes infinite is skipped, and the number of points actually compared is asserted to be non-zero — so a case cannot pass by having had nothing to check.

Test-only change. 5426 pass, 0 fail.

🤖 Generated with Claude Code