[hadashiA]

Fixes #41
Fixes #42

Problem

Non-unique secondary index trees are built from composite keys (source key + 4-byte rid), but TreeBuilder computed entry digests with the source encoding reported by KeyValueList.KeyEncoding, while the read side (NonUniqueSecondaryIndexQuery) digests through DuplicateKeyEncoding, which strips the rid first. The two digest functions disagreed, which broke GetAll/GetRange in two ways:

• 非一意のセカンダリインデックスで、同じインデックスキーが 157 件以上あると GetAll が一部しか返さない #41 (regression in format 1.4): omitKeys was decided from the source encoding's IsKeyDigestExact (true for Int64), so composite-key pages dropped their key bytes even though the digest only covers the source part. Every duplicate collides on the same digest and the rid tiebreaker is unrecoverable, so queries return only part of the entries as soon as the tree spans more than one leaf — hence the 157-row threshold (first leaf split at 156 entries per leaf) and 1000 → 64 (1000 − 6×156).
• Ascii の非一意セカンダリインデックスで、インデックスキーが 8 バイト未満だと GetAll が 1 件しか返さない #42 (present since key digests were introduced): for Ascii source keys shorter than 8 bytes, digesting the composite key leaks the rid's leading byte(s) into the stored digest ("abcdefg" + rid=1 digests as abcdefg\x01). The query digest is computed rid-less, so only the rid-0 entry matches and the bound search stops at the first digest mismatch, returning 1 row. 8-byte-or-longer keys fill the digest before the rid, which is why they were unaffected. The same leak also un-sorts the stored digest array once a key has more than 256 duplicates (rid carry), which this fix covers as well.

Fix

Redefine KeyValueList.KeyEncoding as the encoding of the keys the list actually enumerates: DuplicateKeyValueList overrides it with DuplicateKeyEncoding, which digests the source part only and keeps the default IsKeyDigestExact = false, so TreeBuilder computes digests exactly the way readers do and duplicate-key trees never omit key bytes. The list's internals (Add, the grouping comparer) keep using the source encoding. One file changed; unique trees are unchanged.

Verification

• Both issues' repro code now returns the expected counts (156/157/1000 → all rows; 10 groups × 500 → 500 each; 7-byte and 8-byte Ascii keys → 2 each).
• dotnet test: 95/95 passing.

Note: databases already built with an affected version need to be rebuilt — the bad digests/omitted keys are baked into the file.

🤖 Generated with Claude Code