[ruvnet]

Purpose

Fix the visible ruvector-core::VectorDB::search post-filtering underfill path in current main. Relates to #1310.

Root cause

VectorDB::search retrieved global top-k candidates and then removed disallowed metadata matches without refilling. When the nearest global candidates were from another tenant, the answer could be empty or incomplete despite at least k eligible entries in storage. This matches the failure class documented in the Oct 24 historical comparison; it does not establish byte-for-byte identity with the older npm native package.

Change

• Keep the unfiltered search behavior unchanged.
• Retain the existing fast-path candidate retrieval when at least k results survive the metadata predicate.
• When it underfills, score all matching stored vectors using the configured distance metric, sort with deterministic ID tie-break, and return the eligible top min(k, count).
• Reject NaN/Infinity filtered queries and nonfinite stored vectors/scores during the fallback.
• Add three tests for selective filters (including HNSW) at 1% and 10% among 1,000 vectors, exact ID/order/score/vector matching, fewer than k eligible results, and nonfinite filtered query input.

Explicit boundaries

• This PR changes crates/ruvector-core/src/vector_db.rs, which crates/ruvector-node/src/lib.rs delegates to. It does not update independently shipped npm native binaries, ACORN, FilteredSearch helper, LlamaIndex PR fix(python): enforce LlamaIndex query restrictions and empty-field filters #1150, or duplicate-ID mutation behavior.
• This is a correctness-first fallback, not a filtered-index performance win. It scans N storage IDs and computes distances for M eligible vectors on underfill, potentially incurring large tail latency on disk-backed or million-vector stores. Better indexed prefiltering is follow-up work under Correctness: filtered top-k underfill and vector mutation/input invariants (version-bound) #1310.
• The source patch was inspected against main and regression tests were authored, but Rust/Cargo was not available in the current execution environment. Neither Rust tests nor hosted CI nor before/after benchmarks have been observed to pass. Do not merge until CI and performance review are complete.
• Concurrent writes are not asserted to form an atomic snapshot of metadata and vectors. Authorization is not established by this metadata predicate.

Required validation before merge

cargo test -p ruvector-core filtered_top_k --lib
cargo test -p ruvector-core selective_hnsw --lib
cargo test -p ruvector-core --lib
cargo fmt --all -- --check
cargo clippy -p ruvector-core --all-targets -- -D warnings

Then measure filtered p50/p95/p99 at 10K, 100K and 1M vectors, 1%/10% selectivity, exact oracle and multi-tenant validity. Compare published npm binary separately after rebuilding a pinned native addon.

Acceptance: return min(k, eligible_count), unique permitted IDs, correct score/returned-vector alignment, exact ordering for fallback, and no changed unfiltered behavior.