Add the 1844th performance column, and what two runs of one commit say about the timings

[Rafael-SOWNet]

The performance column 2.4.0 needs, and a measurement that changes how the columns above should
be read.

The Kernel Benchmark was run twice on this same commit

Nothing changed between them.

	run 1	run 2	apart
ParseEasy	12,593 ns	6,846 ns	46%
SimplifyEasy	197,016 ns	111,852 ns	43%
SolveHard	1,638,350,518 ns	887,253,553 ns	46%
RunEasy	50 ns	21 ns	57%
all eighteen			30–57%

Allocation over the same pair: SimplifyEasy 128,098 B against 128,096 B, SolveHard
1,446,783,288 B against 1,446,779,520 B — 0.03%. A thousandfold difference in
reproducibility, which is precisely the argument PerformanceGate already makes for gating
allocation and not time. It had not been measured on this workflow before.

So a timing difference between two columns of this file is evidence only above ~50%. The
1769th's write-up reads −7%, −10%, −15% as improvements and concludes "every row is faster or
flat"; none of those is distinguishable from the runner. Only its SolveEasy row survives —
2.5x with allocation moving 2.3x alongside — and it survives by the rule that write-up itself
states.

The table takes run 2; run 1 is in the prose so the choice is visible rather than quiet.

The exchange cost 13% of SimplifyEasy

Real, and measured where this file cannot measure — one desktop, both arms built and run the
same way, BenchmarkDotNet StdDev under 0.5%:

	SimplifyEasy
1769th, before the rule sets became data	82,676 ns
after the exchange, before the node-type index	148,223 ns
after the index	94,904 ns
1843rd	93,732 ns

Every conversion was measured against the commit in front of it and every one came back free or
better. Nothing was measured against the start. The index recovered most of what the exchange
spent; most is not all.

Counted rather than guessed: 4,914 rule-set invocations at both commits — the work shape is
identical, so this is per-operation overhead, ~1.3 ns per invocation and ~14 ns per rule attempt.
InequalityEquality is 26% of it; InvertNegativePowers, a set with one rule, is 15%. Three
attempts at making dispatch cheaper (one-slot cache, eight-slot cache, array index instead of the
dictionary) were each measured and none moved it. About 40% is still unattributed and this
says so.

The lever, if wanted, is the invocation count rather than the dispatch — 4,914 for one
Simplify, of which 2,202 are the canonical order on every node.

Docs only; no code changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Bjumi5K7fg8yx6UK1mZTQd