Write 12 edge case regression tests and fix the bugs.

[koculu]

All 12 regression tests failed or showed the problem on the old code, and the full suite now passes (844 of 844, 4 m 42 s).

What each fix does

Finding	Test that showed it on the old code	Fix
Rollover can hide data (a new case I found while designing the fix): records with timestamps later than the housekeeping clock ended up past the new boundary	After a rollover, a day's raw query returned 0 instead of 1	Each partition now tracks the latest time written to it, saved in a high-water file at the store root. A new boundary always comes after that day.
Rollover stalls after a clock that ran ahead created an empty future partition	RollOverPartitions returned 0	The partition actually receiving events can now be split, provided its latest written time is known (PartitionedStore.cs).
Salts deleted during a backfill	One anonymous visitor counted 3 times (raw and rollups)	A salt for a past day now records its last use (salt-used: keys) and is destroyed one day after that use. Salts for today and yesterday keep the old date rule (VisitorHasher.cs).
Invalid UTF-16	Rollups counted 2 distinct values where raw counted 1	Text.Name and Text.Value replace unpaired surrogates with U+FFFD before hashing and storing (Text.cs).
Raw codec accepts control characters	Encode didn't throw	RawEventCodec.Encode rejects them; the strict decoder throws FormatException.
SessionState with no dimensions	IndexOutOfRangeException when written, and the session was lost	Missing positions are written as null, meaning "not collected".
Backfill journal size	The journal was 177 MB for 123 MB of rows	It is now a binary file: two alternating generations with a SHA-256 checksum, streamed to disk (journal-a.bin / journal-b.bin).

Existing tests changes

• PublicStorageCodecTests: only the escaping change above. Every assertion is the same.
• EngineTests.cs:663: Updated a comment that described the old salt lifetime. The assertions are unchanged.