Stop Expand losing a term that takes no power

[Rafael-SOWNet]

Closes #851.

The bug

Expand turned RR + 1 into NaN — and with it ZZ + 1, CC + 1, QQ + 1, BB + 1 and true + 1. Because Simplify offers res.Expand() as a candidate and NaN rates as the simplest thing on offer, Simplify returned NaN too.

A 2.0 regression, measured against the published packages:

	1.4.0	2.0.0-preview.1
"RR + 1".Simplify()	RR + 1	NaN
"ZZ + 1".Simplify()	ZZ + 1	NaN

Expand is public, so this is not only a Simplify artefact — "RR + 1".Expand() returned NaN directly.

The cause

CollectLikeTerms is what makes expansion finish: it reduces every factor to a base and an exponent, adds up the terms whose monomials agree, and puts them back together. The reassembly wrote a factor that appeared once as base^1.

Raising to the first power is an identity only where the power is defined at all. It is not, for these:

RR^1 -> NaN     ZZ^1 -> NaN     true^1 -> NaN
{1,2}^1 -> {1,2}   [0;1]^1 -> [0;1]   x^1 -> x   [[1,2],[3,4]]^1 -> [[1,2],[3,4]]

So the decomposition was lossless for everything that takes a power and destroyed everything that does not. That also explains why the sets that can be shifted were never affected: a finite set and an interval both take a power, so they round-tripped.

The fix

A factor that appeared once goes back as itself rather than as base^1.

A second guard covers the same fault above the first power, where the identity does not apply and the factor cannot be written back at all (RR * RR): collecting is an improvement rather than a requirement, so a collection that introduced NaN is discarded and the expanded form kept. That one is deliberately written against the symptom rather than a list of node types, so a node added later cannot reintroduce it.

Measured

Fixed, and the shiftable sets still shift:

RR + 1      Expand: RR + 1      Simplify: RR + 1
true + 1    Expand: True + 1    Simplify: True + 1
{ 1, 2 } + 1                    Simplify: { 2, 3 }
[0; 1] + 1                      Simplify: [1; 2]

Suite: 6061 passed, 0 failed, 14 skipped (11 new cases).

Verified downstream, which is where it came from

Found by running CSharpMath's suite against 2.0.0-preview.1 — the consumer #822 is about. It is 955 green on 1.4.0 and two red on the preview. Rebuilt against this branch:

before: Total: 955, Failed: 2
after:  Total: 955, Failed: 1

The one that remains is 1+\sqrt{2x} expanding to 1+\sqrt{2}\sqrt{x} — the documented 2.0 radical change, sound and theirs to absorb, not a defect here.

Noted, not fixed here

(x+1)^2 * (x+1)^2 expands to ... + 2*x^2 + 4*x^2 + ... with those two left uncollected, which is what CollectLikeTerms exists to prevent. Pre-existing — reproduced on master without this change, so it is not a regression from it and wants its own issue rather than being folded in.

🤖 Generated with Claude Code