(a + b asech(c x))/(d + e x)^2 is integrated, and a root written apart with |x| no longer needs a parity

[Rafael-SOWNet]

Rubi's 7.5.1 (a + b asech(c x))/(d + e x)^2 ran for ten minutes without an answer (the numeric asech(2 + 3x)/x^2 next to it is 0.7 s). Traced with the dispatcher instrumented and a managed stack dump mid-runaway: by parts leaves 1/(x^2 (d + e x) sqrt(1/(c x)^2 − 1)), Euler's substitution answers that in 15 ms, and four things stood between the two.

• A constant over a product is the reciprocal, not the product to the power −1. SolveAsPolynomialTerm's c/g(x) branch handed pe/(x (d + e x) sqrt(…)) on as pe · (x (d + e x) sqrt(…))^(−1), a spelling no radical rule reads (its own comment says so, for the rational rules); it is 1/(…) now, asked as the same question with the constant in front. The 1/g(x) case keeps the power, since 1/g asked again would be the same call.
• SolveByDistributingWholePowersOfProducts asks its respelling as the same question. (c x)^2 and c^2 x^2 are one integrand, and one level down the remainder sat past SolveByWritingAPowerOfAQuotientApart's "question asked or one below" scope — which is what answers it.
• sgn(x)^m where the parity extension fails. The quotient-apart route writes sqrt(1/(c x)^2 − 1) as sqrt(1 − c^2 x^2)/(|c| |x|) and answers for x > 0, then extends by parity — and 1/(x^2 (d + e x) sqrt(…)) has none, so the antiderivative it had just found was declined. |x|^m is sgn(x)^m x^m for a whole m, exact on each half-line, and that stands in for the extension now where every power of x taken is whole (a fractional |x|^(1/2) still needs the parity).
• A root inside a logarithm's argument is left alone. The logarithm is an atom to every rule whatever its argument is written as, and writing asech(c x)'s own root apart re-asked the whole integrand for 1.5 s of nothing.

Measured:

• (a + b asech(c x))/(d + e x)^2 590 s+ → 1.1 s; (a + b asech(c x))/x^7 9 s timeout → 0.09 s; e^(2 asech(a x))/x^2 unevaluated → 0.07 s; x (a + b asech(c x))/(d + e x^2)^2 2.7 → 1.0 s.
• Family 7: 206 → 210 of 270, 0 wrong, 0 error, timeouts 34 → 26, wall 1185 → 952 s (gains: 7.5.1:102/177, 7.5.2:102, 7.6.1:126, 7.3.6:797/883).
• The 1774-problem suite: 1706 → 1707 (Timofeev 944 is answered), 0 wrong, 0 timeout, wall 87 s; per-problem, nothing moved by more than 0.3 s except Timofeev 859, 1 s faster.
• Full suite 12547 passed, 0 failed; the kernel gate PASSED on all 19 gated benchmarks. BREAKING-CHANGES.md entry with the 2.5.0 column measured on a build of v2.5.0 (all four left unevaluated there).

Tests: four rows in InverseHyperbolicSecantIntegralTest, each differentiated back against the integrand at five real points with a, b, c, d, e pinned.

Part of #718.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura