feat(model-vector): dedupe reject-phrase embedding across workers via a required store

[jfallows]

Summary

• VectorModelHandlerImpl is a per-worker factory that eagerly embeds every configured reject phrase in its constructor, once per attach. Since every namespace binding is replicated to and attached independently on every EngineWorker, an N-worker engine fires N identical, redundant embed() calls against the configured embedding provider for the exact same input, all at once, at attach time.
• Add a required store reference on the vector model config (a new StoredConfig, alongside VaultedConfig/GuardedConfig/CatalogedConfig/EmbeddedConfig) so a worker checks a shared cache first, takes a short-lived lock to embed for real only on a cache miss, and every other worker (or, with a distributed store, replica) polls the cache with capped backoff until the winner's result lands. If the lock owner fails without writing a result, the lock simply expires and the next poll wins it instead. A failed embed is never cached, so a transient error doesn't permanently poison the result.
• store is required, not optional: model-vector is incubator-only and has never shipped in a release, so there's no backward-compatibility reason to make the dedup opt-in.

Test plan

• config/engine.conf, incubator/model-vector.conf, incubator/model-vector.spec, incubator/model-vector all green (./mvnw clean verify), plus a full ./mvnw clean install -DskipTests reactor build to confirm nothing else regressed from the new shared StoredConfig class or VectorModelConfig's changed constructor.
• Unit tests in VectorModelHandlerImplTest simulate two workers sharing one store (as store-memory shares its backing map across every EngineWorker in one process): only one of them ever calls EmbeddingHandler.embed(), the loser picks up the winner's cached result via its own retry/backoff, and a failed embed releases the lock for the next attempt without caching the failure.
• VectorModelIT scenarios load a real zilla.yaml referencing a type: test store end-to-end through the engine's own config-loading/name-resolution path, not just direct unit construction.
• SchemaTest confirms the fixture config validates now that store is a required property.