The rational canonical form keeps the points where the expression has no value

[Rafael-SOWNet]

Fixes #1618.

CanonicalizeAsRationalFunction gave an expression a value where it has none: dividing by a quotient moves that quotient's denominator into the numerator, and nothing said the denominator had to be nonzero. On master 6bbed69f against this branch:

….CanonicalizeAsRationalFunction() of	master	this
1/(1/x)	x	x provided not x = 0
1/(1 + 1/x)	x / (x + 1)	x / (x + 1) provided not x = 0
(a/b)/(c/d)	a * d / (b * c)	a * d / (b * c) provided not d = 0
(1/x)^(-1)	x	x provided not x = 0
(x/y)^(-2)	y ^ 2 / x ^ 2	y ^ 2 / x ^ 2 provided not y = 0
x/(y/x)	x ^ 2 / y	x ^ 2 / y provided not x = 0
0/x	0	0 provided not x = 0
x^2/x^2	1 provided not x ^ 2 = 0	1 provided not x = 0
(x^2 - 1)/(x^2 + 2x + 1) + 1/(x + 1)	x / (x + 1) provided not x ^ 2 + 2 * x + 1 = 0	x / (x + 1)

What changes. The form gathers through the step AsSingleFraction takes (#1622), SingleQuotient.OverLeastCommonDenominator, which collects each denominator it turns over, so the two agree on what a single quotient of an expression is and on where it has a value. Those denominators are kept as a condition beside the factors a cancellation removes, and so is the denominator under a numerator that vanishes. The condition is part of a form whose point is that equal trees mean equal functions, so it is written canonically too: the square-free part of what was excluded -- E / gcd(E, dE/dx_1, …, dE/dx_n) -- without the factors the reduced denominator still excludes. So x^2/x^2 and x/x meet, and the last row says nothing the denominator does not already say. A sum widens nothing: it is defined exactly where its terms are, and the common denominator vanishes exactly where one of theirs does.

Tests: RationalCanonicalFormTest -- 1/(1/x) meets x^2/x, x^2/x^2 meets x/x and (a/b)/(c/d) meets a d^2/(b c d); 1/(1/x), 1/(1 + 1/x) and 0/x no longer collide with x, x/(x + 1) and 0; seven expressions evaluated at the point where they have no value, form and expression both NaN; and the denominator's exclusion not repeated. Each of the new rows fails on master.

Measured. The suite passes, 14,760 tests with 13 skipped, on this change over #1622's last head; cut from master after #1622 merged, the canonical-form, single-fraction, transformation and corpus tests pass, 928 of them. Simplify reaches the form inside the interval-membership rule, so the corpus gate is among them. BREAKING-CHANGES.md has the entry, both columns measured on a build, v2.5.0 against this change.

The allocation gate run here failed on EvalPolynomialFresh, 13,672 against 12,536 B/op, with the machine at a load of thirteen and the benchmark at 28 times its baseline time. That benchmark substitutes a decimal into x^3 + 2 x^2 - x + 1 and evaluates it, which this change does not reach, and measured directly -- the same call twenty thousand times after a warm-up -- it allocates 12,536 B/op on this branch, on master 6bbed69f and on 8f3757cd alike. CI's gate is the one to read.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura