Every registered rule set runs as data (#825)

[Rafael-SOWNet]

The three CanonicalOrder sets were the last to run their switch. MatchedRules.Sort(level) had
existed as their data form, held to Patterns.SortRules(level) by MatchedRulesAgreeWithTheSwitchTest
at a hundred firings per level — and been run by nothing. They are repointed at it.

30 of 30 registered sets now execute the matcher, and 0 take their Rules from
RuleRegistryGenerator. Both were 27 and 29 when RuleMetadataTest's remark was written; it says
so now.

Measured, not inferred from the agreement test

8,310 generated inputs simplified on a build of master and on this branch, output diffed: zero
differences. The agreement test says the two spellings agree rule-by-rule; this says Simplify
agrees at the entry point, which is the thing a caller sees.

And it is not free

The write-up of the last rule-set exchange in this repository ends with "a run of individually-free
steps needs one measurement against where it started, or it has not been measured at all." So,
allocation — master (eb3a0447) against this change, one benchmark on one machine:

benchmark	master	this
SimplifyHard	3,655,695,192	3,662,375,184	+0.18%
SolveHard	1,450,058,856	1,452,704,384	+0.18%
SolveMediumHard	165,055,344	165,611,184	+0.34%
SimplifyEasy	128,099	128,460	+0.28%
CompileHard	20,753	20,434	−1.5% — inside that family's ±2% floor, means nothing

Real, because those Simplify/Solve rows reproduce to one part in 10⁵. Inside PerformanceGate's
3%. And the third small step since 2.4.0 on this family, after #1176 and #1177 — roughly half a
per cent cumulative. The Sort set runs on every node of every pass, so a matcher dispatch where a
switch used to be is the likeliest source; it is the price of the set being addressable, named and
tiered like the other twenty-seven. Stated so it is a known cost rather than a surprise later.

(The A/B's first table came out empty because I restored key-commits.txt while the script was still
running — it reads the list a second time at combine time. The measurements were on disk; only the
table had to be redrawn. Worth knowing before the next person restores that file mid-run.)

The gate tripped, correctly

RuleMetadataTest.EveryRuleOfARepointedSetCarriesItsIdentity requires that a set is repointed only
once every rule of it carries an identity, so the registry gains descriptions rather than trading
them. The nine Sort rules had none — they rendered as Xorf x => (built by code). I had repointed
a set that wasn't ready. The fix is at the source: each of the nine carries a description now.
The gate is untouched.

Five pins moved because the exchange succeeded

Each to the number its failure reported:

pin	was	is
sets describing what they run	27	30
addressable rules	315	321
rules carrying an identity (two tests, and WritingARule.md)	294	321

One pin was wrong while passing. "How many registered sets run the matcher" matched
MatchedRules names — and CanonicalOrder runs under the name Sort, so it said 27 while the truth
was 30, for exactly as long as that family went unlisted. It has the same special case
CommonDenominator already had.

StepAsASentenceTest asserted that some set was still generator-described and that its names did
not pass for prose. That population is now empty, and the assertion flips to say so rather than
being deleted: a set falling back to the generator is a regression this should catch.

On the "six orphan sets"

The audit that prompted this named six MatchedRules sets as orphans. Measured: three are the
CommonDenominator family, wired all along. The other three — PowerOfPower, PythagoreanIdentity,
SharedFactor — are single-rule fixtures that ReversibleRuleTest, GatheredMatchingTest and the
e-match tests use as subjects; deleting them deletes the evidence. Two duplicate registry rules under
near-identical names. SharedFactor's a-shared-factor-comes-out-of-a-sum appears nowhere else,
which is worth its own look and is not taken here.

Part of #825 and #746.

Checks

Full suite 9571 passed, 0 failed, 14 skipped. 8,310-input sweep: 0 changed answers.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura