bug: align mem_wal HNSW construction with the reference algorithm

Summary

rust/lance/src/dataset/mem_wal/hnsw/graph.rs implements a third, independent HNSW construction path, alongside the offline HnswBuilder and the OnlineHnswBuilder in lance-index. It did not benefit from the reference-alignment audit in #8036, which #8188 applied to the other two builders.

This issue tracks applying the same reference alignment to the mem_wal builder so that all three construction paths share one tested set of invariants.

Deviations to audit and fix (mirroring #8036 / #8188)

1. New-node degree limit. max_neighbors(m, level) returns 2 * m at level 0 for all pruning, including a newly inserted node's initial neighbor selection. Per Algorithm 1 and hnswlib, a new node selects M neighbors; Mmax0 = 2M only bounds reciprocal edges on existing level-0 nodes.
2. Reciprocal edges vs. diversity pruning. Verify that reciprocal candidates join the complete old-plus-new candidate set before SELECT-NEIGHBORS, with no distance-only cutoff gate (the failure mode #8188 removed from the other two builders).
3. Algorithm 4 configuration. select_neighbors should document and implement the chosen options — extendCandidates = false, keepPrunedConnections = true (refill from pruned candidates up to the degree limit), matching #8188. The refill is load-bearing: correcting the new-node limit to M without refill fragmented the graph enough to make HNSW-PQ recall fluctuate stochastically (0.902–0.987) in #8188's CI.
4. Entry point / level assignment. mem_wal already uses random levels plus entry-point promotion; verify the semantics match the reference (the first globally highest node is the fixpoint of sequential strict promotion) and that to_lance_hnsw_batch stays consistent with the serialized level_offsets contract (configured level count, empty trailing ranges) expected by lance-index's HNSW::load.
5. Construction parameter validation. mem_wal uses its own BuildParams; apply the same preconditions as HnswBuildParams::validate() from #8188 (max_level > 0, m >= 4, ef_construction >= m, overflow guard) at its boundary, with descriptive errors.

Acceptance criteria

• Per-role degree limits (M for a new node, Mmax0 for reciprocal capacity) with structural tests.
• Reciprocal candidates participate in the diversity heuristic with the complete candidate set.
• keepPrunedConnections refill implemented and Algorithm 4 options documented.
• Entry-point semantics verified against the reference model.
• Parameter validation with descriptive errors and tests.
• Existing mem_wal HNSW tests and the HNSW::load round-trip test keep passing; recall does not regress.

References

• #8036 — umbrella audit of HNSW deviations from the reference algorithm
• #8188 — reference alignment for the offline and online lance-index builders
• #5156, #5183 — related root causes fixed there

Start with rust/lance/src/dataset/mem_wal/hnsw/graph.rs and compare its construction logic with the reference-alignment changes in #8036 and #8188. Review HnswBuildParams::validate() and the HNSW::load level_offsets contract, then run the existing mem_wal HNSW tests and load round-trip test. Done means the listed degree, pruning, entry-point, validation, and recall criteria pass.