PR1: data-driven voice-action registry (Telegram as first built-in)

[initcore0]

First step of generalizing named voice actions. Replaces the hardcoded enum Action { case telegramPost } + actionPhrases table + single editable telegramPostPrompt with a data-driven registry, so built-in and pack/user actions share one mechanism — and packs can later add new actions (tweet, commit message…), not only retune the Telegram prompt.

This is behavior-preserving; it only adds the extensibility seam PR2 (packs) will use.

Changes

• VoiceAction {id, displayName, triggerPhrases, prompt} + VoiceActionRegistry (allPhrases, action(id:), merging(_:) = override-by-id / append-new) in OpenWhispCore, Codable so actions can travel in a pack later. The built-in Telegram action carries the exact phrases + default prompt that used to live in AppState.
• VoiceCommandParser: Result.action: Action? → actionID: String?; matches an injected registry (default .builtins), phrases tried longest-first so a specific phrase beats a substring. Wake / imperative / generic paths unchanged.
• AppState: applyVoiceCommand looks up the action by id and uses its prompt; new computed voiceActionRegistry overlays the user's editable telegramPostPrompt onto the builtins. defaultTelegramPostPrompt becomes an alias.
• Adversarial review confirmed the Telegram + generic paths are behavior-identical (all 12 EN/RU phrases preserved, default prompt byte-identical, LLM directive identical in empty/default/custom cases). The only difference is a cosmetic status string ("Making Telegram post…" → "Running Telegram post…").
• Tests: +10 (registry merge/override/codable, custom action triggers by id, longest-phrase-wins); parser tests migrated to actionID. 181 → 191.

swift test 191/191 · ./build.sh clean.

🤖 Generated with Claude Code