Make the Instruction agent reply in the user's language, at a human pace

The agent answers in the language each message is written in (Indonesian
or English), whatever the UI language; short replies such as "LLM" keep
the conversation's language. inLang() runs t() in that language.

Each turn now reads like a person: a thinking step with what the agent is
doing, the reply typed out with "Thought for Ns" above it, a short typing
pause, then one question, with its options shown once it is typed.
Reduced motion skips the animation.

"dari file LLM" / "from the document X" is read as the topic, and topics
keep the user's capitals.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

src/express/agent.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it } from 'vitest'
-import { EMPTY, nextQuestion, respond, type AgentState } from './agent'
+import { chatLangOf, EMPTY, nextQuestion, respond, type AgentState } from './agent'
 import { t } from '../i18n'
 
 const say = (...inputs: string[]) => inputs.reduce<AgentState>((s, input) => respond(s, input).state, EMPTY)
@@ -71,4 +71,16 @@ describe('instruction agent', () => {
     expect(nextQuestion(s)?.type).toBe('tooMany')
     expect(rules(say('10 summaries per day about sales', '30 per week'))[0]).toMatchObject({ count: 30, per: 'week' })
   })
+
+  it('takes "dari file X" as the topic', () => {
+    expect(rules(say('Saya mau buat video dari file LLM'))[0]).toMatchObject({ kind: 'video', topic: 'LLM' })
+  })
+})
+
+describe('chatLangOf', () => {
+  it('tells Indonesian from English, and stays undecided on short replies', () => {
+    expect(chatLangOf('Saya mau buat video dari file LLM')).toBe('id')
+    expect(chatLangOf('I want to make 2 quizzes per week')).toBe('en')
+    expect(chatLangOf('LLM')).toBeNull()
+  })
 })