Speeds up rule evaluation on large libraries, and fixes a recent Sonarr/Radarr lookup regression. No new dependencies.

1. Parallelize per-item operand resolution

Rules resolve each item's operand one at a time; for watch-history-style properties (no bulk endpoint, #2936) and per-item *arr lookups this dominates runtime. The comparator now resolves firstVal/secondVal in bounded parallel batches (chunk + Promise.all — the same idiom as the existing Plex collection-write batching), then runs the unchanged backward mutation loop.

• RULE_EVALUATION_CONCURRENCY = 8 — the single global cap and the only batching layer (never nested inside getters), abort checked inside each task.
• Conservative by design: the binding constraint is the slowest co-located backend (e.g. Tautulli's CPU-heavy history queries), not host core count. Higher values over-drive such backends past their request timeout, which then retries and amplifies load.

2. Fix the #2897 Sonarr/Radarr lookup regression

#2897 made getSeriesByTvdbId/getMovieByTmdbId uncached so the empty-show cleanup reads post-deletion truth (#2757 / #2891). Correct for the cleanup — but it also de-cached the rule-evaluation path, where the same series/movie resolves once per item, so episode-level rules over large libraries went from a handful of cached lookups to one uncached call per item per section.

ArrLookupCache restores the de-duplication without reintroducing the bug: a run-scoped memo created for the evaluation loop only and never handed to the collection/action phase, so the cleanup still reads fresh and a stale entry can't survive a deletion. The API-layer lookups stay uncached; only the two uncached *arr identity lookups use it (Tautulli/Seerr/Plex/Jellyfin/Emby already cache at the API layer).

Real-world result

~13k-episode Tautulli+Sonarr rule on an all-in-one Intel N100:

8 is faster than 16 because removing the backend overload eliminated the timeout/retry stalls.

Tests: ArrLookupCache unit tests, an action-handler test asserting the cleanup reads series fresh (never via the memo), and the existing #2897 regression suites all pass.