[Rafael-SOWNet]

IntervalEvaluation.Agree took the size of each value as sqrt(re^2 + im^2) in doubles. Past about 1.3e154 the square is infinite, so the tolerance relative to it was infinite too, and any two values that large agreed. Two kinds of check went through it:

• Coefficients read in lowest terms at pinned symbols (IsZeroAtPinnedSymbols, reached through InLowestTermsOverTheSymbols and IsZeroAsAValue): a coefficient that large at the pins was taken for zero.
• Sampled-point checks (HoldsAtSampledPoints, DerivativeHoldsAtSampledPoints): any answer that large at a sampled point passed, whatever it was.

A wrong answer on master that 2.5.0 declined, so it never shipped: (b x^(-4) + b x^(-6) + b x^(-8))/sqrt(1/c^2 + x^2). Its antiderivative's coefficients run to c^294, about 2e167 at the pin c = 3.71. Two of its three terms were taken for zero and dropped, and the answer was wrong at every point. (b x^12 + b x^10 + b x^8)/(x^16 sqrt(1/c^2 + x^2)) was the same. Both are answered correctly now. This came to light as a wrong answer in a pending branch that asked this question in exactly this spelling.

The size is now the hypotenuse scaled by its larger part, so it never squares past the doubles. The bounds are widened outward, and a scale that is still not finite is left to the decimals (null).

Measured on 8a1ef454, branch against master:

Run	Problems	Master	Branch
Pocket: slow and timed-out problems, and Rubi's 7.5, 7.6 and 1.2.1.9	1,208	1,019 solved	1,020 solved
Family 1, 40 a file	1,381	1,362	1,362
Family 0, whole	1,814	1,789	1,789
Families 2 to 8, sampled	2,410	2,362	2,356

Re-run alone on both builds, every move is a problem near the budget. Two that master answered in 22–23 s timed out on the branch (7.3.5:50, 7.4.1:369); both have flipped either way in other runs today. No run had a wrong answer.

IntervalEvaluationTest gains a case where two values past the doubles must not agree, and one for a large complex size. ACoefficientPastTheDoublesIsNotTakenForZeroIntegralTest differentiates both antiderivatives back; it fails on master. BREAKING-CHANGES.md has the entry.

Part of #1019.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura