tweak(mi-voice): brief but well-formed, not telegraphic

Prior voice-mode prompt leaned too far into caveman/telegraphic speech
('drop articles'). User wants brevity with proper grammar — short
complete sentences, not neanderthal fragments.

server/mi-routes.ts

@@ -306,14 +306,14 @@ Results will be provided in a follow-up message for you to incorporate into your
 [MI_ACTION:{"type":"batch","actions":[...actions...],"requireConfirm":true}]
 Use requireConfirm:true for destructive batches.${voiceMode ? `
 
-## VOICE MODE — SPEAK LIKE A CAVEMAN
-User is hearing your reply read aloud. Be extremely terse.
-- Max 1-2 short sentences per reply. Fragments fine. Drop articles (a/an/the) and filler.
-- No lists, no headers, no markdown, no emoji, no code blocks — plain spoken prose only.
-- No preamble ("Sure!", "Of course,", "Great question"). Answer first, stop.
-- Never narrate what you're about to do. Do it. Report result in ≤1 sentence.
-- Action markers still allowed — they are silent and do not count toward length.
-- If a full answer would take longer than ~8 seconds to read aloud, give the headline only and offer to expand.` : ""}`;
+## VOICE MODE — BRIEF BUT WELL-FORMED
+User is hearing your reply read aloud. Be brief, but speak in complete, well-formed sentences.
+- Max 1–2 short sentences per reply. Normal grammar, normal articles, normal English — no telegraphic or caveman speak.
+- No lists, no headers, no markdown, no emoji, no code blocks — natural spoken prose only.
+- No preamble ("Sure!", "Of course,", "Great question"). Answer first, then stop.
+- Don't narrate what you're about to do. Do it, then state the result in one sentence.
+- Action markers still allowed — they are silent and don't count toward length.
+- If a full answer would take more than ~8 seconds to read aloud, give the headline only and offer to expand.` : ""}`;
 
 	// Build conversation
 	const miMessages: MiMessage[] = [