Expand a cancelled quotient of factorials instead of throwing (#817)

[Rafael-SOWNet]

Closes #817.

What was wrong

Entity.Expand cancels a quotient of factorials before it does anything else, and the cancellation can leave a sum — (x + 1)! / x! is x + 1. The result went straight back into SmartExpandOver, which is documented <summary>expr is NEITHER Sumf NOR Minusf</summary> and asserts that it never receives one. So the assertion fired and an AngouriBugException — the library reporting that the library is broken — came out of a public method, on an expression Simplify answered as 1 + x throughout.

case Divf:
    expr = expr.Replace(Patterns.ExpandFactorialDivisions);   // (x + 1)! / x!  ->  x + 1
    if (!(expr is Divf(var dividend, var divisor)))
        return SmartExpandOver(expr, ...);                    // ... which must not take a sum

Thrown rather than returned, so a caller mapping Expand over a corpus loses the whole run instead of one entry.

Why this fix is right

GatherLinearChildrenOverSumAndExpand is the sum-aware entry point, and it already delegates to SmartExpandOver for anything that is not a sum. Calling it here is one identifier and a strict superset of the old behaviour: identical where the cancelled expression is not a sum, and correct instead of fatal where it is.

The other two recursive calls are unaffected — one is handed a product, and the public MathS.SmartExpandOver splits by Sumf.LinearChildren before descending, which is exactly the guard that was missing here.

Measured

On this build, each of these threw AngouriBugException before:

input	now
(x + 1)! / x!	x + 1
(x + 2)! / x!	x ^ 2 + 3 * x + 2
(x + 3)! / x!	x ^ 3 + 6 * x ^ 2 + 11 * x + 6
x! / (x + 1)!	1 / (x + 1)
(x + 1)! / x! + y	x + 1 + y
((x + 1)! / x!) * (a + b)	a * x + b * x + a + b
(x + y + 1)! / (x + y)!	x + y + 1

Checked as values, not strings, at a point where the factorials are defined: 5!/4! = 5, 6!/4! = 30, 7!/4! = 210 — and the expanded form agrees with the original at that point in every case.

sin(x)! / x! is unchanged and still comes back as written, which is correct: nothing cancels.

Tests

5808 passing, 0 failed, 14 skipped — the skip this issue was holding is gone. F# 130 passing.

Two regression tests: one asserting the value is preserved, one asserting no factorial survives where the two cancel completely. Both were checked against the unfixed code — five cases fail without the change, including the nested (+ y) and multiplied (* (a + b)) ones, so the tests are not passing vacuously.

Compatibility

BREAKING-CHANGES.md records it: the same call returns a value where it used to throw. Nothing could have depended on the old behaviour — AngouriBugException is not a contract — but the entry belongs there anyway since the answer changed.

🤖 Generated with Claude Code