fix(cache): give each cache instance its own dict

[JediBrooker]

What

_cache was declared as a class attribute on both SimpleCache and StringConfigCache, so every instance in the app shared a single dict.

class SimpleCache[VT, *KTs]:
    _cache: dict[tuple[*KTs], tuple[int, VT]] = {}   # shared by every instance

For StringConfigCache this is harmless today — the config classes have disjoint key literals, so they were only sharing storage.

For SimpleCache it's a bug. prowlarr_source_cache and prowlarr_indexer_cache both key off a single string, so cached search results land in the same keyspace as cached indexers. get_indexers() then reads those search results back out:

indexers = list(prowlarr_indexer_cache.get_all(source_ttl).values())
if len(indexers) > 0:
    return IndexerResponse(indexers={indexer.id: indexer for indexer in indexers}, ...)

indexer is a list[ProwlarrSource], and the whole thing lands in the except Exception in get_indexers.

Reproducing

Search for any book, then open Settings > Indexers. It renders the error state with:

Failed to access Prowlarr to fetch indexers  error="'list' object has no attribute 'id'"

Minimal version:

from datetime import datetime
from app.internal.models import TorrentSource
from app.internal.prowlarr.util import prowlarr_source_cache, prowlarr_indexer_cache

src = TorrentSource(guid="g1", indexer_id=1, indexer="MAM", title="Some Release", size=1,
                    publish_date=datetime.now(), info_url=None, indexer_flags=[],
                    seeders=1, leechers=0)
prowlarr_source_cache.set([src], "Dune")

# the indexer cache now returns the search results
assert prowlarr_indexer_cache.get_all(9999) == {}   # fails on main

It's intermittent in practice, which is probably why it hasn't been pinned down: it only shows up once something has been cached, and flush_prowlarr_cache() (called when the Prowlarr settings are saved) happens to paper over it — flush() does self._cache = {}, which creates an instance attribute and silently diverges from the shared class one from then on. So the fix also makes flush() mean what it looks like it means.

The change

Moves both dicts into __init__. Six lines, no behaviour change beyond the instances no longer colliding.

Checked

• uv run basedpyright — clean
• uv run ruff format --check app — clean
• uv run alembic check on a fresh DB — no new operations (no schema change here)
• every SimpleCache/StringConfigCache subclass is instantiated exactly once as a module-level singleton, so nothing depended on the sharing
• config key literals are disjoint across AuthConfig, oidcConfig, ProwlarrConfig, ABSConfig and the quality/indexer configs, so no config loses a value it was previously reading out of a sibling's cache

Written with AI assistance; the diagnosis and the reproduction above were verified by hand against main.