[bghira]

After the main process refreshes Webshart buckets, the data-backend factory calls reload_cache() on each rank. That method only reloaded the bucket index, leaving workers with stale or empty caption and image-metadata caches. Caption requests could then fall back to per-sample backend lookups instead of using the cache already written by the main process.

Load image metadata and captions inside WebshartMetadataBackend.reload_cache(), and remove the redundant explicit loads from initialization and bucket discovery. Add a regression test using separate backend instances sharing real serialized caches. It checks both set_config values, replacement of stale captions and metadata, restored bucket membership, and caption access without backend lookups.

Validation:

• Confirmed the new regression test fails against the implementation on main for both set_config values because it calls the caption backend.
• .venv/bin/python -m unittest -v -f tests.test_webshart_backend: all 40 tests passed.
• .venv/bin/python -m unittest -v -f: 7,059 tests run, 286 skipped, no failures, using LyCORIS 4.1.3.
• Formatting, import sorting, and lint hooks passed.