feat: Tauri backend commands + wikipedia rewrite + github PR review pack

GUI integration
  - crates/jarvis-gui/src/tauri_commands/backends.rs — 5 new Tauri commands:
      get_active_backends()  -> {tts, llm, llm_model, profile}
      set_llm_backend(name)  -> swap LLM hot, persists to DB
      set_tts_backend(name)  -> persist choice (effective on jarvis-app restart)
      llm_reset_context()    -> clear conversation history
      llm_get_last_reply()   -> last assistant message or null
  - Registered in main.rs invoke_handler. Svelte side can now read active
    backend + swap from a settings page / footer without env-var gymnastics.

Wikipedia pack rewritten (resources/commands/wikipedia/)
  - Old `wiki/wiki.lua` (104 lines, hand-rolled Groq HTTP + inline SAPI PS) deleted.
  - Replacement `wikipedia/summary.lua` (87 lines):
    * REST summary endpoint (cleaner than opensearch+summary chain).
    * Russian first, English fallback with LLM translation.
    * Pure LLM fallback if Wikipedia has no article.
    * Uses jarvis.cmd.{ok,error,not_found} + jarvis.llm + jarvis.speak (no
      inline PowerShell SAPI). Phrases kept compatible with old pack.

GitHub PR voice review (resources/commands/github_pr/, 3 commands)
  - "текущий репо bossiara13/J.A.R.V.I.S-rust"  → github.set_repo
    Persists repo via jarvis.memory.
  - "какие пиары" / "открытые пиары"            → github.list_prs
    Calls `gh pr list --json number,title,author,createdAt --limit 10`.
    Parses count + first 3 titles by regex from JSON.
  - "разбери последний пиар" / "что в pr"       → github.summarize_pr
    `gh pr view N --json title,body,additions,deletions,changedFiles,author`,
    sends to LLM with "senior reviewer" system prompt, speaks 3-5 sentence
    review focusing on changes, risks, merge-readiness.
  - Requires gh CLI installed and `gh auth login` done.

Tests: 52/52 jarvis-core unit tests pass (Wikipedia pack TOML/script auto-checked).
Build: cargo build --release -p jarvis-app -p jarvis-gui green.

resources/commands/wikipedia/command.toml

@@ -0,0 +1,40 @@
+# Wikipedia summary — fetch a topic's intro from Wikipedia REST API, speak it.
+
+[[commands]]
+id = "wikipedia.summary"
+type = "lua"
+script = "summary.lua"
+sandbox = "full"
+timeout = 15000
+
+[commands.phrases]
+ru = [
+    "что такое",
+    "кто такой",
+    "кто такая",
+    "расскажи про",
+    "википедия",
+    "найди в википедии",
+    "что говорит википедия",
+    "поищи в википедии",
+    "википедия о",
+    "википедия про",
+    "найди в википедии про",
+    "статья из википедии",
+]
+en = [
+    "wikipedia",
+    "look up in wikipedia",
+    "wikipedia article on",
+    "what is",
+    "who is",
+    "tell me about",
+    "wikipedia about",
+]
+ua = [
+    "вікіпедія",
+    "знайди у вікіпедії",
+    "що таке",
+    "хто такий",
+    "розкажи про",
+]

resources/commands/wikipedia/summary.lua

@@ -0,0 +1,105 @@
+-- "Википедия Эйнштейн" → fetch ru.wikipedia.org/api/rest_v1/page/summary/<topic>
+-- Falls back to LLM if Wikipedia doesn't have the article.
+local phrase = (jarvis.context.phrase or "")
+
+local query = jarvis.text.strip_trigger(phrase:lower(), {
+    "найди в википедии про",
+    "найди в википедии",
+    "что говорит википедия",
+    "поищи в википедии",
+    "статья из википедии",
+    "википедия про",
+    "википедия о",
+    "википедия",
+    "расскажи про",
+    "что такое",
+    "кто такая",
+    "кто такой",
+    "look up in wikipedia",
+    "wikipedia article on",
+    "wikipedia about",
+    "wikipedia",
+    "tell me about",
+    "what is",
+    "who is",
+    "знайди у вікіпедії",
+    "вікіпедія",
+    "що таке",
+    "хто такий",
+    "розкажи про",
+})
+query = query:gsub("^[%s,:%.]+", ""):gsub("%s+$", "")
+
+if query == "" then
+    return jarvis.cmd.error("Какая статья?")
+end
+
+-- Wikipedia title needs underscores for spaces. Capitalize first letter of each word
+-- to match canonical titles (good enough heuristic).
+local title = query:gsub("(%S+)", function(w)
+    -- Capitalize first byte (works for ASCII; for cyrillic, leave alone).
+    local first = w:sub(1, 1)
+    if first:match("[%a]") then
+        return first:upper() .. w:sub(2)
+    end
+    return w
+end):gsub(" ", "_")
+
+local function urlencode(s)
+    return (s:gsub("[^A-Za-z0-9%-_%.~]", function(c)
+        return string.format("%%%02X", string.byte(c))
+    end))
+end
+
+-- Try Russian first, fall back to English.
+local lang_chain = { "ru", "en" }
+local data, source_lang
+
+for _, lang in ipairs(lang_chain) do
+    local url = string.format("https://%s.wikipedia.org/api/rest_v1/page/summary/%s",
+        lang, urlencode(title))
+    local d = jarvis.http.json(url)
+    if d and d.extract and d.extract ~= "" then
+        data = d
+        source_lang = lang
+        break
+    end
+end
+
+if not data then
+    -- last resort: ask LLM directly
+    local llm_answer = jarvis.llm({
+        { role = "system", content = "Кратко (2-3 предложения) на русском объясни тему. Если не знаешь — скажи." },
+        { role = "user", content = "Расскажи про: " .. query }
+    }, { max_tokens = 200, temperature = 0.3 })
+    if llm_answer and llm_answer ~= "" then
+        return jarvis.cmd.ok(llm_answer)
+    end
+    return jarvis.cmd.not_found("В Википедии нет статьи про " .. query .. ".")
+end
+
+local extract = data.extract
+local title_actual = data.title or query
+
+-- Truncate extract to ~3 sentences for spoken output
+local sentences = {}
+for sentence in extract:gmatch("[^%.%!%?]+[%.%!%?]") do
+    table.insert(sentences, sentence)
+    if #sentences >= 3 then break end
+end
+local spoken = #sentences > 0 and table.concat(sentences) or extract:sub(1, 400)
+
+if source_lang == "en" then
+    -- Translate English extract to Russian via LLM (no point in speaking English
+    -- through a Russian voice).
+    local translated = jarvis.llm({
+        { role = "system", content = "Переведи английский текст на русский, без вводных и комментариев." },
+        { role = "user", content = spoken }
+    }, { max_tokens = 300, temperature = 0.0 })
+    if translated and translated ~= "" then
+        spoken = translated
+    end
+end
+
+jarvis.system.notify("Wikipedia: " .. title_actual, spoken:sub(1, 200))
+return jarvis.cmd.ok(spoken)