v0.2.0 — smarter memory + security 🛡️

Feature: memory that finds things by meaning and keeps facts current, plus deterministic hardening against prompt injection and secret leaks.

• 🧠 Search by meaning, not the exact word — recall used to be a raw grep: type the wrong synonym or a different word-form and the fact stayed hidden. Memory is now a ranked search — BM25 over your cards and summaries (built in on node:sqlite, no dependency, no index to babysit) plus a rerank by the links between cards, so asking about a person surfaces the people and projects connected to them. It's language-agnostic by design: term weight comes from your own vault, so it works the same in Russian, English, Chinese, or a mix — no hardcoded word lists.
• ♻️ Facts that change get rewritten, not stacked — when today contradicts an old card (you changed jobs, moved city, a decision was reverted), the nightly rollup now rewrites the current value and files the old one under a dated ## History, instead of leaving two contradictory facts for search to trip over. A deterministic nightly pass flags same-entity conflicts; each fact is tagged EXTRACTED (you said it) or INFERRED (deduced) so Iva can hedge when it's guessing. The longer you use it, the fewer stale contradictions it carries.
• 🔒 Strictly typed cards — a new write_card path validates type and schema at write time, so the model can't invent a card type or smuggle in stray fields; the nightly maintenance coerces anything written the old way back into schema.
• 🛡️ Untrusted content is gated — forwarded messages, attachments, voice transcripts and web pages now pass through deterministic gates before and after the model: hidden prompt-injection (invisible characters, homoglyph tricks, override phrases) is defused on the way in, and API keys, tokens and exfiltration URLs are scrubbed from replies on the way out. Legitimate non-Latin text is never mangled — normalization is used only to detect, never to rewrite what you wrote.
• 🔌 Optional semantic search — for a large vault or genuinely fuzzy/cross-language queries, an opt-in hybrid mode adds vector search fused with the keyword ranking (RRF). One external key (Jina — no-train/EU, or DeepInfra — cheapest), or point it at a local Ollama endpoint and use no external key at all. Off by default; the installer asks once. Base memory needs nothing.