Sweep every span query, mutating ones included

[mattlorimor]

TestSpanAndArrayQueriesAgree reached 5 of the 20 structures offering a span query. It now reaches all 19 with a pure one, and a second sweep covers the 13 whose span query also changes the structure.

Two properties, not one

The mutating half needed a different shape. Agreeing on every answer is only half of it: two paths can return the same value at every step and leave the structure holding different things. A filter that answers correctly while holding the wrong thing doesn't fail at the seam — it fails later, somewhere else, for no visible reason.

So each pair in TestSpanAndArrayMutatingQueriesAgree is driven step for step through its own instance, comparing the answer at every key, and then compared through its payload — the same oracle the equivalence sweep uses.

Two mutators (InvertibleBloomLookupTable.Remove, SublimeCountMinSketch.Remove) answer with the structure itself, so there is no value to compare and the state is the whole check. The table also needs the fixed-width helper, since it refuses a key of any other length.

Rosters

Both sweeps check against rosters derived from the library's span surface, split by whether the method mutates:

• WithPureSpanQueries — 19 structures (Test, Count, TryGetValue)
• WithMutatingSpanQueries — 13 (TestAndAdd, TestAndRemove, Remove)

A structure that gains a span query fails these until it is covered.

Verification

Five probes, each with the build exit checked:

• InvertibleBloomLookupTable.Remove's array path applying twice → killed with left different states. This is the one that matters: a fluent mutator has no answer to compare, so nothing but the state check could see it.
• BloomFilter.TestAndAdd's array path negating its result → killed with answered differently for key 0. The mirror case: identical state, wrong answer.
• dropping InfiniFilter from the mutating roster → killed, named
• dropping HeavyKeeper from the pure roster → killed, named
• the pre-existing equivalence and allocation sweeps still pass

Clean dotnet build -c Release -warnaserror, exit 0 unpiped. 907 tests passing, up 1.

A comment I had to correct

I first annotated PrivateCountMinSketch and DpswSketch as drawing noise per query, which would mean their two overloads need only agree in distribution. That's wrong, and the tests passing is what exposed it: the noise is drawn once at construction and lives in the counters — which is precisely what stops repeated queries from averaging it away. A query is an ordinary read, and the two overloads must return the identical double. The comment now says that.