Name the next release 2.1.0, with the performance pair and the #746 tier note

[Rafael-SOWNet]

Release preparation for 2.1.0, carrying the two checklist items 2.0.0 shipped without.

The version

Version and FileVersion go to 2.1.0; AssemblyVersion stays 2.0.0.0, which is what the comment
beside it asks for — the assembly is strong-named, so every AssemblyVersion change breaks binding for a
consumer without a redirect, and holding it makes each 2.x a drop-in replacement.

Not a patch. The sixteen commits since the tag change answers, withdraw two deliberately, and reserve a
word (NaN).

The performance pair

v2.0.0 (the 1691st commit) against the candidate (the 1707th), 18 benchmarks each, measured minutes
apart on one machine in one session — because columns from different hardware cannot be read as a ratio,
which is what the top of that file has always said.

No regression. The largest real move is SolveMedium at +4.6%; the rest sits inside ±3%. Allocation
is flat, and Derivate, EvalTrig and EvalTrigPrecise are byte-identical.

It corrects the pair above it

That one reported SimplifyEasy at +8.7%, attributed it to guards calling Evaled inside a pattern's
when clause, and asked for "a measurement of its own before more guards are written that way".

Six more guards of exactly that shape have landed since — #887's interval tests, #892's zero test, #881's
sign read, #902's positivity check, #897's sort guard. SimplifyEasy is +2.1% here, and its
allocation moves 0.5%. If the guards were the cause, the row should have moved further and allocation
should have followed, since Evaled allocates on its first call per node. Neither happened.

So the +8.7% was mostly the machine, the attribution in that section was wrong, and the action item it
raised is discharged — the pattern is not measurably expensive at this scale. Recorded in the file, since
the wrong reading is as instructive as the number.

The changelog said the wrong thing about which version shipped what

All thirteen entries added after the tag sat under ## 2.0.0 — since 1.4.0, which tells a reader they are
in the package they already have. They now sit under ## 2.1.0 — since 2.0.0 with their own glance table.

Three of them I had misremembered as pre-tag — the Compile exception (#894), the logarithm division
(#890) and the arctan/arccotan guard (#887) — which is why the split was derived by diffing the file
against v2.0.0 rather than judged by eye. Section and row counts were checked before and after: 20
rows and 13 sections moved, nothing duplicated, nothing dropped. My first attempt at the move did
duplicate content — 76 sections became 192, because the span of each section ran to the next ## and
swallowed the ### sections between — which the count caught and a backup undid.

The #746 tier note

Posted as a comment on #746, and it says
plainly that 2.1.0 advances no tier. Against v1.0's required infrastructure: the polynomial layer
untouched, the canonical-form specification untouched, pattern-matching-as-data untouched — this release
added switch arms — and assumptions-travelling-with-a-node untouched. The one item that moved is the
corpus-and-harness requirement.

Of the three standing conditions, two are met (performance measured; correctness coverage grown) and one
is unaffected (no package boundary moved).

The note also carries an argument for the roadmap: three separate pieces of work this session stopped at
the same wall — #902's two lost limits, #721/#890's domain contradiction, and the ln(a) + ln(b) guard
— and each is a guard-versus-coverage trade that exists only because an assumption cannot be attached to
an expression and carried. Three independent hits is a stronger case for scheduling v1.0's metadata item
early than any one of them is alone.

Checklist

Suite and harnesses green, output read	✔ 6488 C# + 130 F#; casbench 117/119 with 0 wrong; rootcheck 596/596; simpsweep 10463/10463; propcheck 1340 with 0 failures; crashcheck 1652 with 0 crashes; boundcheck 2 disagreements, both awaiting the #721 decision; docsamples 84 samples with 0 mismatches; sympycheck 24 programs with 0 failures
BREAKING-CHANGES.md entry per changed answer, measured on a build of each	✔ and now filed under the right version
Performance column, previous re-measured on the same machine, published as a pair	✔
Version checked against #746, note saying which tier it does and does not advance	✔
Wiki and website quickstart still true	✔ docsamples clean; the Solvers page was updated when #901 shipped

What I have not done: tagged, packed or published anything. This is the preparation; the release itself
is yours.