fix(udf): l2_distance returns L2sq to match USearch and DuckDB

[anoop-narang]

Summary

• Remove .sqrt() from l2_kernel so l2_distance UDF returns squared L2, matching USearch's MetricKind::L2sq and DuckDB VSS's array_distance
• Previously the UDF returned actual L2 (with sqrt) while all rewritten paths returned L2sq, causing the same query to produce different numeric distance values depending on whether the optimizer fired
• Update README distance metrics documentation

Test plan

• cargo fmt --check
• cargo clippy -- -D warnings
• cargo test — all 32 tests pass (no tests assert on specific distance values, only ordering)

[claude]

One stale doc comment not covered by the diff: src/lib.rs:96 still says — Euclidean distance (L2). Should be updated to — squared L2 distance (L2sq, no sqrt) to match the new semantic. Non-blocking.

[claude]

P2 (non-blocking): No comment here explaining that the sqrt omission is intentional. A future maintainer may add .sqrt() back thinking it's a missing step, silently reintroducing the inconsistency. Consider:

Suggested change

	}
	// Returns L2sq (no sqrt) — matches USearch MetricKind::L2sq and keeps numeric
	// values consistent between the UDF path and the optimizer-rewritten index path.
	}

[claude]

Review

Issues

P1 — No numeric regression test (tests/execution.rs)

The entire motivation for this PR is fixing a numeric inconsistency (UDF returned L2, index returned L2sq). However, the test suite only asserts on ordering, not on actual distance values. This means the fix has zero regression coverage: if someone adds .sqrt() back to l2_kernel, every test will still pass.

A minimal test would cover the gap — e.g., using a non-trivial query where L2 ≠ L2sq:

// [3,4,0,0] vs [0,0,0,0] → L2sq = 25.0, L2 = 5.0
// Assert the returned value is 25.0, not 5.0
SELECT l2_distance(vector, ARRAY[3.0::float, 4.0::float, 0.0::float, 0.0::float]) AS d
FROM items WHERE id = 3

Without this, the consistency guarantee introduced by the fix is not tested.

Action Required

Add at least one test that asserts a specific numeric value for l2_distance so the squared-L2 semantics are regression-protected.