perf: revisit a Set-based innermost index layer (follow-up to #418)

[jeswr]

Tracking issue recording the conditions for picking the #418 experiment back up, per its close-out.

#418 replaced the innermost layer of the three-layer quad index with Sets. Its own numbers showed a clear insert-path regression (load 23.3s → 39.7s, 855 MB → 1083 MB), and read-path benchmarks were never produced to offset it. Anyone revisiting this should start from a fresh branch and:

• benchmark read-heavy workloads (match/getQuads/countQuads over large stores) alongside the insert path, using the perf/ scripts;
• adapt the index-level set operations that have landed since (perf: compute intersection over indices #438, perf: perform matching over indices #443, perf: perform difference using internals #451, perf: selectively perform matches #458, perf: merge indices directly for cross-entity-index set operations #601) — they walk the index with for...in / key in and assume plain-object leaves, so Set leaves would silently break intersection/difference today (git reports no conflicts, which makes this easy to miss);
• mind Node compatibility: Set.prototype.isSubsetOf needs Node >= 22 while engines declares >= 12, so set-algebra fast paths need feature detection or another shape;
• sequence against the entity-index work in memory: make entity index global and use a WeakMap #434, which may reshape the same layer.