test(match): zsync sparse-match adversarial fixture

[oferchen]

Summary

Adds crates/match/tests/sparse_match_fixture.rs, an integration test that drives the production DeltaSignatureIndex::from_signature + DeltaGenerator::generate pipeline over synthetic basis/source pairs where almost no basis blocks recur in source. Exercises the rolling-hash hot path under near-constant rejection, the worst case identified in docs/design/zsync-inspired-matching.md that the planned bithash prefilter (#2059) targets. Pins today's pre-bithash matching accuracy as a baseline so future prefilter work can be benchmarked without regression risk.

Matrix coverage

• Planted matching block counts K in {0, 1, 2} (K=0 is the most important: full literal stream, exercises the no-match hot path).
• Basis and source sizes in {64 KB, 1 MB, 16 MB}. The 16 MB cases are marked #[ignore] so default cargo nextest runs only the 64 KB and 1 MB fixtures; stress-test CI matrices opt in via --run-ignored only.
• Block sizes in {1024, 4096}.
• Strong checksums: MD5 (no seed) and XXH3-64.
• Worst-case timing budget of 5s asserted on the 64 KB / 1024-block fixture as a tripwire for accidental quadratic regressions.

Wire-compat invariants confirmed

• The pipeline emits exactly K Copy tokens at basis indices 0..K, in order.
• matched_bytes == K * block_size, literal_bytes == source.len() - K * block_size, total_bytes == source.len().
• Byte-domain disjointness between basis (in [0, 0x7f]) and source non-planted bytes (in [0x80, 0xff]) makes the "no-match" property robust at every sliding-window alignment, not just block-aligned offsets. A construction sanity-check asserts this directly on the smallest fixture.
• Strong-checksum verify call counts are not directly observable through the public API; the test notes this in a comment with the upstream reference (rsum.c:362-366) and asserts only the byte-level invariants.

Test plan

• CI: fmt+clippy, nextest (stable), Windows, macOS, Linux musl all green.
• Default cargo nextest run -p matching skips the #[ignore]-marked 16 MB cases.
• cargo nextest run -p matching --run-ignored only runs the stress fixtures in the opt-in matrix.