Read a rule set that is a single rule (#825) — re-delivering #969 to master

[Rafael-SOWNet]

This re-delivers #969, which merged but did not reach master. My mistake, and worth stating plainly: I opened #969 stacked on feat/addressable-rule-factories (#968). #968 merged to master first, so when #969 was merged it went into its base branch rather than master — the squash commit a38d455a sits on feat/addressable-rule-factories, and master has none of it:

$ git show origin/master:…/RewriteRules.cs | grep -c 'InvertNegativePowersArms\|PolynomialLongDivisionArms'
0

I should have retargeted #969's base the moment #968 landed. Nothing is wrong with the code — it is the same two commits, cherry-picked onto current master (upstream squash-merges, so a rebase would have replayed #968's commits against themselves).

The content, unchanged from #969:

Three rule sets are one rule each, written the way a single rule is written when there is no need for a switch:

x is pattern [&& guard] ? replacement : x

InvertNegativePowers, PolynomialLongDivision, PolynomialGcdCancellation. The generator reads that shape and emits exactly one rule — pattern from the is, guard from whatever conjuncts follow it, replacement from the true branch.

• The false branch has to be the subject itself. Anything else is a second rewrite, and emitting one rule for it would silently drop the other.
• The condition is kept whole rather than rebuilt from pattern + guard, because a later conjunct may declare what the replacement reads: TreeAnalyzer.PolynomialLongDivision(num, denom) is var (divided, remainder) is exactly that, and divided + remainder needs it.

	before	after
addressable sets	19	22
addressable rules	368	371
armless sets	11	8

The second commit corrects a claim I made in the first: that what remains is armless "because it is a method with branches and locals". Only three of the eight are. CollapseMultipleFractions is an ordinary one-parameter expr switch, PerfectSquare is this PR's own shape, and the three CommonDenominator sets are a switch with a second parameter — so the list is a list, not a category. A follow-up takes four of those five.

The tests bind

RewriteRules.PolynomialGcdCancellation  ->  Patterns.PolynomialLongDivisionArms

  Failed AddressableRulesTest.EveryArmIsTheArmItWasGeneratedFrom(setName: "PolynomialGcdCancellation")
  Failed: 1, Passed: 28

Reverted; 29/29 pass. Two defects found by reading the generated output rather than assuming it: <see cref="X"/> left See for what is and is not attempted with a hole in it, and tag-stripping applied to // comments ate a ^ (-1) => 1 / a down to a ^ (-1) 1 / a. Only /// is XML now — verified across all 371 rules, 92 described, 8 containing < or >, all 8 intact.

Suite 7304 passed, 0 failed, 14 skipped, corpus gate included. No answer changes: ApplyOnce is handed the same Func, only the optional rule list beside it is new.