Refuse an absurd payload for every structure, not just the newest

[mattlorimor]

TestPersistenceHostilePayloads exercised the field-level guards of 18 of the 33 structures. The drift ran the opposite way from the sweeps fixed in #132 and #133: the structures added most recently were the best covered, and fifteen of the oldest had no field-level test at all — HyperLogLog, DDSketch, ThetaSketch, CountMinSketch, QuotientFilter, BinaryFuseFilter, BloomierFilter, CountSketch, SimHashSignature, InvertibleBloomLookupTable, HyperLogLogPlus, and the four older Bloom variants.

Their readers already had the guards. Nothing exercised them, so a guard that had quietly stopped firing would have been found by a user with an edited file rather than here.

What this covers

Eighteen new tests, each poking one field, repairing the CRC, and asserting the message its own guard produces — the message is the landing assertion, which is what keeps a payload from being refused for the wrong reason. That mattered: three of my fragments ("between", "hash function", "at least one") were loose enough to match an unrelated refusal, and I tightened them and re-ran to confirm the tests still landed where intended.

The offsets come from FORMAT.md, which #133 made complete and testable.

A map, not a loop

TestEveryStructureRefusesSomeAbsurdPayload is the completeness check. It can't be one loop the way the persistence sweeps are, because what counts as absurd differs per structure. So it's a map from each structure to the test carrying it, with both halves checked: the roster comes from StructureId, and each named test must exist and carry [TestMethod].

I first wrote that map with a DeclaredElsewhere exemption list claiming fifteen structures were covered in their own test files. That was false — they were covered in this same file all along, and my exemptions would have quietly excused them from the roster. The map is now all 33 entries with no exemptions, generated by parsing the file rather than transcribed.

Verification

• Removing HyperLogLog's reader guard → its test fails
• Removing BinaryFuseFilter's fingerprint-width guard → its test fails
• Widening ThetaSketch's trailing-value tolerance to every value → its test fails
• Adding a member to StructureId → the completeness check fails by name
• Stripping [TestMethod] from a test the map vouches for → fails naming the test

My first attempt at the HyperLogLog probe reported SURVIVED, which I chased before trusting it: the surgery had disabled the constructor's power-of-two check rather than the reader's, so the test was never going to notice. Retargeted at the InvalidDataException site, it kills. A probe that lands in the wrong place is indistinguishable from a test that doesn't work, until you look.

Also removes an orphaned doc comment that had drifted off PokeUInt64 and stacked a second <summary> onto PokeByte.

Clean dotnet build -c Release -warnaserror, exit 0 unpiped. 925 tests passing, up 19.