Seven rules say how much they expand (#825)

[Rafael-SOWNet]

Growth is Unknown for a code replacement unless it is declared, and 262 rules sat there. Seven can
say what they do. 255 now do not.

Argued from the code, measured against the corpus

Two mechanisms, the same arithmetic.

Six negation rules — (-a) + (-b), (-a) - x, and the four that bring a negative factor out of
a product, a numerator or a denominator. The operators map one for one, the literal is replaced by
its magnitude, and the negation the replacement puts round the whole of it is two nodes:

y / (-a * x)   ->   -(y / (a * x))          five nodes to seven, for every input

One power rule — a ^ n = 1 / a ^ (-n). The power stays a power with its exponent replaced one
for one; the quotient and its numerator 1 are the two.

Each was also run against the generated corpus by RuleGrowthAgreesWithTheCorpusTest, between 9 and
84 firings apiece, none contradicting. That is corroboration and not the argument — the corpus can
only refute.

Why Expands is the right first batch

Not only because it is what they do. Saturation.RulesUpTo builds its set as the rules up to
Rearranges, so:

• declaring Collects or Rearranges adds a rule to what equality saturation fires;
• declaring Expands changes nothing about what runs.

These seven record something that was already true without moving anything. The Collects and
Rearranges candidates in the measurement are real too, but each of those is a behavioural change
and wants the saturation measured before and after — a separate PR, not a line in this one.

The eighth, left alone

a-negated-reciprocal-rational-factor-is-a-negated-division measures +2 on all 18 of its
firings, which is exactly the evidence the other seven have. It stays Unknown.

Working through its pattern I count five nodes to five, so the measurement and my reading of the
code disagree, and I cannot argue the claim. WritingARule.md says that where you cannot argue it
from the code you leave it Unknown — and that sentence was written in the previous PR, by me, so
it applies here first.

Declaring it on corpus agreement alone would be the same mistake #1158 caught: a growth asserted
from what it looked like rather than from what it does.

Verification

Full suite 9552 passed, 0 failed, 14 skipped. The census moves 262 → 255 in
WritingARule.md and RuleAuthoringGuideTest.

Part of #825.