Fix EliasFanoVec::rank returning out-of-range index for last element of large bucket

[rustamch]

Summary

EliasFanoVec::rank(v) can return a value >= len() for a value v that is actually present in the vector, when v is the last element of an EF upper-bucket whose size exceeds BIN_SEARCH_THRESHOLD (= 4).

The bug is a single-line off-by-start_index_lower in the binary-search fallback of search_element_in_block::<INDEX=true, UPWARD=true> at elias_fano/mod.rs:386:

if INDEX {
    cursor = final_bound as isize + direction;  // ABSOLUTE
}
break;
// ... falls through to:
return if INDEX {
    start_index_lower as u64 + cursor as u64    // treats cursor as RELATIVE
}

final_bound is an absolute index into lower_vec, but the fallthrough return adds start_index_lower to cursor, double-counting it.

Discovery

Found via property testing in a downstream crate that stores two parallel EliasFanoVecs (term IDs and a per-term auxiliary value) and looks up the auxiliary via vec_b.get_unchecked(vec_a.rank(found)). For certain inputs, the bogus rank exceeds vec_b.len() and causes an out-of-bounds panic at bit_vec/mod.rs:1187 (get_bits_unchecked's data[pos / WORD_SIZE]).

A minimal counterexample (shrunk from 368 → 14 elements) is pinned as test_rank_last_element_of_large_bucket. For the 14-element vector, indices 5..9 all fall into upper-bucket 4; rank(term_at_idx_9) returns 14 (= len) before the fix and 9 after.

Fix

Subtract start_index_lower from cursor when reassigning it in the fallthrough path so the final return remains correct:

if INDEX {
    cursor = final_bound as isize + direction - start_index_lower as isize;
}

Verification

• New regression test test_rank_last_element_of_large_bucket (fails before, passes after).
• Full test suite: 199 unit + 72 doc tests pass (cargo test and cargo test --release).

Notes

Only search_element_in_block's INDEX=true fallthrough branch is affected. successor, predecessor, and get_unchecked — and rank for inputs whose bucket size is ≤ BIN_SEARCH_THRESHOLD — are unaffected. No public API changes.

[rustamch]

Thanks, for merging fix so quickly!