A root the substitution search would write through i is kept, and the cube is answered too

[Rafael-SOWNet]

Closes #1370: sqrt(a + b sech(x)) tanh(x)^5 and sqrt(a + b sech(x)) tanh(x)^3 are both answered now, and right.

The tangent substitution writes sech(x) as sqrt(1 - u^2) for u = tanh(x). The substitution search then takes u_2 = u^2 - 1 and simplifies the rewritten integrand one level, and that writes b sqrt(-u_2) as i b sqrt(u_2). That is the other root wherever u_2 is negative, and tanh(x)^2 - 1 is negative everywhere. The fifth power was answered in i sqrt(u^2 - 1), and its derivative was off at every real point, until the tangent substitution's check turned it into a decline.

Two commits:

• The first passes over a candidate whose integrand holds the imaginary unit where the integrand did not, and another candidate answers the fifth power.
• The second keeps a candidate's integrand as written where the one-level simplification would bring in i, at each of the four places the search simplifies one. Traced, the cube's simplification went one step further than the fifth power's, from sqrt(-u_2) to i sqrt(u_2), so the candidate that answers it was passed over. With the root as written, the search in u_2 answers the cube through sqrt(-u_2), as it answers the fifth power. The first commit's check stays, for whichever step writes i otherwise.

Input	Was (2.5.0)	Now
sqrt(a+b*sech(x))*tanh(x)^5	ran past two minutes	the antiderivative, in sqrt(a + b sqrt(1 - tanh(x)^2)), written in exponentials
sqrt(a+b*sech(x))*tanh(x)^3	ran past five minutes	the antiderivative, the same way

ExponentialSubstitutionIntegralTest.ARootOfTheSecantBesideAnOddPowerOfTheTangentIsAnswered differentiates both back at five points with the symbols pinned.

Measured at 780f638, both commits, against master 40b7e58:

• The whole of Rubi 4.1.12, 4.2.12, 6.5.3 and 6.5.7 (555 problems that count, 3 s each): master 455 solved, this branch 461, 0 wrong. The six that moved are all of 6.5.3's odd powers of tanh(c + d x) beside sqrt(a + b sech(c + d x)), its reciprocal and its minus three-halves power, the fifth powers and the cubes. Re-run alone, they are this branch's and no other PR's. The files' total time went from 470 s to 416 s.
• A sample of families 1 to 7 (5 per file, 912 problems): 764 solved on both, 0 wrong, row for row.
• The independent suites (family 0, 1814 problems, 5 s each): 1750 on both, 0 wrong, 0 timeouts.
• The unit tests pass (net10.0, 13327 passed), every target builds, and the allocation gate passes at 780f638: allocation is what the baseline says on all 19 gated benchmarks.

🤖 Generated with Claude Code