v3.24.0

Added

• Sentinel finding explanations can now be written in your language — a new optional sentinel_response_language setting in the Sentinel subentry (next to the explanation toggle), contributed by @hruba202 (#523). Set it to e.g. Czech and the LLM-authored explanation shown in mobile and persistent notifications is generated in that language; leave it empty (the default) for English. The option is deliberately scoped to the explainer only: discovery candidate titles/summaries and triage decisions are parsed by code downstream (normalization, dedup, evidence matching) and always stay in English — #524 tracks making discovery prose translatable safely. Mirrors vlm_response_language's convention.
• Privacy redaction stays deterministic under translation — inflected languages decline names by grammatical case ("Petra" becomes "Petru"), which would let a translated explanation slip a recognized person's name past the notifier's exact-match redaction. For sensitive findings with a response language set, recognized names are now removed from the model's input before the explanation is generated — walking the entire evidence structure (values, keys, captions, nested containers), matching longest-first so overlapping names like "Alex"/"Alexander" can't leave a partial name behind. A name the model never sees cannot be emitted in any inflection; the prompt's keep-names-nominative instruction remains as defense in depth. English behavior is unchanged. Found and iterated through cross-model adversarial review.