[tisonkun]

Simplify the unreleased XOR API to XorFilter::from_values, XorFilter::from_hashes, and XorFilter::from_parts. from_values hashes and deduplicates ordinary values; from_hashes accepts distinct hashes as &[u64] and returns an error for duplicates. Fingerprint width and construction seed are explicit arguments.

Keep serialization beside the filter implementation so fields and serialization helpers remain private. Remove the builder, raw-parts wrapper, and ownership-transfer method; application-defined codecs import raw fields and encode through borrowed accessors. Use concise local names where the surrounding operation already supplies the context.

Update documentation and existing coverage for value deduplication, invalid hash inputs, bounded retries, Java serialization, and xorf interoperability. Repeated-hash benchmarks now include explicit caller-side deduplication for both implementations.

Validation: cargo x check, cargo x test (814 tests including doctests), cargo x lint, and the optimized XOR benchmark smoke suite with ThinLTO.

Batch construction at e3b508e, Apple M4 Max / macOS ARM64, nightly-2026-08-28, ThinLTO, three repetitions. Each round builds 32 independent Xor8 filters from identical distinct inputs with initial seed 1. Times include setup and cleanup, reported as median [min–max] milliseconds:

Keys per filter	Rounds	DataSketches	xorf 0.13.0
128	10,000	271 [266–280]	417 [410–420]
16,384	40	142 [140–142]	359 [357–365]
50,000	12	167 [167–168]	337 [337–338]

Command: CARGO_PROFILE_BENCH_LTO=thin cargo +nightly-2026-08-28 bench -p benchmarks --bench benchmarks -- xor::build_batches --repeat 3 --raw. These compare the current implementations; the API simplification does not change the construction algorithm.