RVF 0.3.4 does not persist INDEX_SEG; reopened 28k-vector stores take >60s on first query

Summary

The current public packages (@ruvector/rvf@0.3.4, @ruvector/rvf-node@0.2.3, both current latest on 2026-07-30) do not emit a persisted index / INDEX_SEG when an RVF store is created and populated through the documented RvfDatabase.create() + ingestBatch() path.

This makes reopened large stores pay an extreme first-query cost. In RuvNet Brain's real source corpus:

• 1,466 vectors / 4.4 MB: first query ~5.5s, second query 3-15ms
• 13,764 vectors / 43 MB: first query exceeded 60s under the measured run
• 28,152 vectors / 83 MB: first query exceeded 60s under the measured run

The same stores open in 9-105ms. The delay occurs on the first db.query().

Segment proof

db.segments() on the 28,152-vector store reports 330 segments containing only:

• manifest
• vec
• witness
• journal

There is no index segment. The 1,466-vector store likewise has 23 segments and no index.

A controlled 100-vector reproduction also emits no index:

const db = await RvfDatabase.create(file, {
  dimensions: 4,
  metric: 'cosine',
  m: 16,
  efConstruction: 200,
  overwrite: true,
});
await db.ingestBatch(rows);
console.log(await db.segments());
await db.compact();
console.log(await db.segments());

Before compact: manifest, vec, witness, manifest.
After compact: vec, witness, witness.
Still no index.

Expected

The documented HNSW creation options (m, efConstruction) should produce a durable INDEX_SEG, and openReadonly() should use it so first-query latency is near steady-state query latency rather than scaling into tens of seconds.

If the index is intentionally embedded inside VEC_SEG, please document that layout and investigate why the first query scales this way after reopen. The public README currently identifies 0x02 INDEX_SEG as the HNSW graph segment.

Impact

This blocks an interactive <4s search SLO and caused shared MCP worker timeouts at 30-240 seconds. Raising the timeout is not an acceptable workaround.

Start with the documented RvfDatabase.create(), ingestBatch(), openReadonly(), segments(), and compact() paths, then reproduce the 100-vector case and inspect the README's INDEX_SEG description. Done means the documented creation path produces a durable index used after reopen, or the VEC_SEG layout is documented and first-query latency no longer scales into the reported timeouts.