[Rafael-SOWNet]

Fixes #2022.

lim x->-oo ln(x) + ln(x + 1) - ln(x^2 + x) was 0, on 2.5.0 and on master. Both x and x + 1 are negative there, so each of the first two logarithms carries i pi and the limit is 2 pi i.

The cause. On an approach the limit machinery gathers ln(a) + ln(b) into ln(a b) where a and b are both positive, and it reads their signs from the approach in scope. SolveByRewriting reads -oo as +oo with -x written for x, then simplified the result while the approach in scope still said -oo. Read against -oo, -x and 1 - x are positive, so ln(-x) + ln(1 - x) was gathered and cancelled against ln(x^2 - x). The rewriting now runs with +oo in scope, as the Gruntz branch and SimplifyAndComputeLimitToInfinity already did.

The memo. #2015's nested-Simplify memo is keyed on the tree, the level and the settings, and not on the approach. On an approach a simplification can gather logarithms it does not gather on its own, so one tree has an answer per approach. The memo is now bypassed while an approach is being read. Through Simplify of a limit(...) node, master answered 2 ln(-1) only because the memo carried an answer across approaches; the bypass alone gives 0, and the bypass with the limit fix gives 2 ln(-1).

Measured on builds of v2.5.0, master 6e16aae and this branch:

….Limit("x", "-oo") of	v2.5.0	master	this
ln(x) + ln(x + 1) - ln(x^2 + x)	0	0	ln(-1) + ln(-1)
ln(x) + ln(x - 1) - ln(x^2 - x)	0	0	ln(-1) + ln(-1)
(ln(x) + ln(x + 1) - ln(x^2 + x)) x/(x + 1)	0	0	ln(-1) + ln(-1)
(ln(x) + ln(x + 1) - ln(x^2 + x))^2	0	0	(ln(-1) + ln(-1))^2

LogarithmsGatheredOnTheirOwnApproachTest holds these four and six limits that were already right, through both Limit and Simplify. Without the SolveByRewriting change the four fail and the six pass. The suite passes locally: 16353 passed, 0 failed. Indefinite integration never reads an approach, so the bypass does not reach the integrator.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura