v0.5.0

Changed

• Log levels reclassified so an embedding application is no longer flooded by default.
  Nitrite logged 253 statements at error!, and 204 of them sat immediately before the matching
  Err(...) return — every rejected filter, every failed type conversion, every "collection name
  cannot be empty" validation was reported twice: once through the returned NitriteError, which
  the caller handles, and once screamed into the host application's log, which the caller cannot
  suppress without silencing Nitrite entirely. On a normal workload the second copy is pure noise,
  and it arrives at the one level every application leaves enabled.

  Levels now follow a single rule — log at a level proportional to what the caller cannot already
  see:

  • debug — the error is returned to the caller as Err. The NitriteError carries the same
    message and kind, so the caller decides the severity. 245 statements moved here.
  • warn — an anomaly Nitrite absorbed and the caller never learns about: a skipped
    corrupt catalog/index entry, a corrupt index (whose key detail exists only in the log line, as
    the propagated error is a shared static), best-effort background compaction/GC/flush failures,
    an invalid regex that silently degrades a filter to "matches nothing", a scheduled task
    dropped because its duration could not be converted.
  • error — a swallowed failure that may have left state inconsistent: the four rollback paths
    in batch insert/update, and close/drop failures for the index manager, plugin manager, Fjall
    map/store, and keyspace drain. Ten statements remain at this level, down from 253.

  A default info-level application therefore sees only failures it can act on. Full detail is
  still one target filter away and unchanged in content:
  RUST_LOG=warn,nitrite=debug,nitrite_fjall_adapter=debug.

  Two other level changes fall out of the same rule: Snowflake no longer announces its node id
  at info! on every generator construction (now debug!), and nitrite_vector's "index was
  stale or damaged; rebuilt from collection" moved from info! to warn!, matching the
  already-warn! sibling messages in the HNSW and DiskANN loaders that report the same class of
  damage.

  No message text, error type, ErrorKind, or control flow changed — this is purely which
  statements reach a default logger.