Read the quotient in the limit instead of gathering it in the simplifier (#802)

[Rafael-SOWNet]

Fixes #802 — the quotient twin of #801, and without the cost I said it would have.

a^n / b^n was gathered into (a/b)^n unconditionally, which is false across the branch cuts: sqrt(2)/sqrt(-3) is -0.8165i where (2/-3)^(1/2) is +0.8165i.

I left this out of #801 because guarding it removed limit answers — (x^2 + 1)^x / (x^2)^x stopped having a limit at all, which is the whole of #739 and #740. That framing was wrong, and this is the correction.

The gathering was never the point

#740 wanted it so the limit machinery could read a 1^oo out of a quotient — its own test comment says exactly that. So the rewrite belongs in the limit reader, not the simplifier, because that is the only place it can be justified:

• a limit needs the identity to hold only in a neighbourhood of the destination, so requiring both bases to be eventually positive there is enough — and there is a destination to check that against;
• in the simplifier there is no destination and nothing to check, which is precisely why the rule was unconditional.

ApplySecondRemarkable now recognises a^n / b^n itself, and the simplifier's rule takes the same guard the rest of its family carries.

Every limit survives

(x^2 + 1)^x / (x^2)^x	1	unchanged
(x^3 + 1)^x / (x^3)^x	1	unchanged
(x - 5)^x / x^x	e^(-5)	unchanged
(sqrt(x) + 1)^x / sqrt(x)^x		unchanged
sqrt(x) / sqrt(y)	was sqrt(x/y)	unchanged, and correct at x=2, y=-3

Eleven tests moved from the mechanism to the outcome

PowerQuotientGatheringTest asserted that Simplify prints a single power. It no longer does, and should not — those assertions were pinning something unsound. They now assert that the limit is answered, which is what the mechanism existed for. The four cases with symbolic bases no longer gather at all, and assert instead that simplifying does not change their value, since nothing can say a symbolic quotient stays on the principal branch.

The test that recorded #802 as open is flipped to assert it stays fixed — which is what it was written to do.

Harnesses

• unit 5543 pass, 0 fail; F# 130/130
• casbench 113/117, 0 wrong; rootcheck 596/596; simpsweep 10463/10463; propcheck 0 failures

🤖 Generated with Claude Code