The re-embed's index DDL runs outside the transaction, not inside it

[jasperblues]

Fixes a deadlock introduced by #116. Found by changing the embedding model on a live appliance, not by reading code.

#116 put the index drop and remake inside reembedAll. This class carries a class-level @Transactional, so the whole method ran in one transaction — and Neo4j will not take schema work and data work together. The DDL runs on its own session and waits on the schema locks the open data transaction holds, while that transaction waits to commit.

It does not fail. It hangs.

Observed changing text-embedding-3-small → text-embedding-3-large on a real appliance:

Reindex: text-embedding-3-small (dim=1536) -> text-embedding-3-large (dim=3072)
Index drift on Proposition[embedding]: existing dimensions=1536 ... requested dimensions=3072

…then ten minutes of silence, the HTTP request still open, and Neo4j reporting:

neo4j-transaction-20296   status=Running   elapsed=PT10M27S   currentQuery=NULL

The graph was left with the proposition index dropped and the vectors not rewritten — the worst of both, and precisely the inconsistency #116 set out to prevent. Chunks had already moved to 3072, because DrivineStore.reembedAll is not transactional.

Fix

The DDL sits either side of the transaction; only the rewrite is transactional:

@Transactional(propagation = Propagation.NOT_SUPPORTED)
override fun reembedAll(): PropositionReembedReport {
    val shapeChanged = persistenceManager.indexes.ensure(spec) is EnsureResult.Drift
    if (shapeChanged) persistenceManager.indexes.drop(spec)
    val rewritten = rewriteEveryEmbedding()          // @Transactional
    if (shapeChanged) persistenceManager.indexes.ensure(spec)
    …
}

NOT_SUPPORTED rather than simply removing the annotation: the class default would otherwise still apply, and it additionally suspends a transaction a caller opened — the same hazard arriving from outside. save() already overrides the class default for its own reasons and is the precedent.

Why the existing test passed — the part worth keeping

It constructs the repository directly, following the precedent of the other tests in this file:

DrivinePropositionRepository(graphObjectManager, persistenceManager, FakeEmbeddingService(width), transactionManager)

No Spring proxy, so no transaction, so the DDL and the writes landed in separate transactions. It passed on a path production never takes. It exercised the class; the bean is what production calls.

The new test

Uses the autowired bean, and forces the shape mismatch on the index rather than needing a second model. A deadlock doesn't fail, it hangs — so it is wrapped in assertTimeoutPreemptively, which turns a stuck build into a failure.

Verified both ways: with @Transactional put back, it fails with execution timed out after 60000 ms. With the fix, both re-embed tests pass.