Add key digest arrays to BTree nodes (format 1.1)

[hadashiA]

Summary

B+Tree nodes now carry a contiguous array of 8-byte order-preserving key digests, and searches probe that array instead of dereferencing the variable-length keys scattered across the page. One cache line covers eight probes, and the per-probe two-step dependent load (meta → key) collapses to a single load.

Case	Before	After
Point lookup, 10k rows (fixed key)	24.8 µs	18.1 µs	-27%
Point lookup, 10k rows (random keys)	43.4 µs	33.4 µs	-23%
Point lookup, 1M rows (fixed key)	59.7 µs	53.9 µs	-10%
Point lookup, 1M rows (random keys)	88.2 µs	87.7 µs	neutral
Range scan, 1M rows	45.9 µs	47.8 µs	neutral

(1 op = 1000 queries. The 1M cases are DRAM-latency bound, where fewer touched lines and a ~20% larger page count cancel out; the win is in the CPU/cache-bound regime.)

Digest definitions

IKeyEncoding gains SupportsKeyDigest / GetKeyDigest (default interface members, so custom encodings are unaffected):

• Int64LittleEndian — the value with its sign bit flipped: an exact total order, so searches never touch key bytes at all.
• Ascii — first 8 key bytes packed big-endian (zero-padded): byte-lexicographic order maps onto unsigned comparison; keys sharing an 8-byte prefix collide and fall back to the full comparison.
• DuplicateKeyEncoding (non-unique secondary indexes) — delegates to the source encoding's digest of the key without the rid suffix.
• UUIDv7 / custom encodings — no digest (Guid.CompareTo is not byte-lexicographic); their trees keep the plain layout automatically.

Format & compatibility

• Page layout (digest pages): [PageHeader][NodeHeader][digest u64 × N][meta × N][payload].
• The flag lives in the upper bits of the node header's kind field (NodeFlags.HasKeyDigests), per page — digest and plain trees coexist in one file, and 1.0 files remain fully readable.
• DatabaseBuilder.KeyDigests (default true) — disable to produce byte-identical 1.0 output (header stays MinorVersion 0) at the cost of the search speedup. File-size cost when enabled: +8 bytes/entry (~+27% for this benchmark's 13-byte values; proportionally less for larger values).
• Files built with digests are marked MinorVersion 1. Note: published 1.0 readers do not validate the version and cannot parse digest pages — worth a release-note callout.

Also

• Fixed a latent bound-search bug (exists on main): Search/SearchAsync with LowerBound/UpperBound returned not-found when the bound was the first entry of the next leaf; they now continue to the right sibling. The new page layout shifted leaf boundaries and exposed this in existing range tests.
• New benchmark: BigReadBenchmark (1M rows, ~40 MB) covering the DRAM-bound regime.
• New tests: digest-collision fallback (keys sharing an 8-byte prefix) and the KeyDigests = false layout.

All 75 tests pass.

🤖 Generated with Claude Code