feat: TTS usage v1 — input character fallback (openai legacy, google, kotodama)

Part of #1415. Follow-up from #1420 research. ## What Surface input-character usage for TTS providers whose upstream response carries no per-call usage. Pure derivation from `namedInputs.text` — no SDK rework. Targets: - `src/agents/tts_openai_agent.ts` for `tts-1` and `tts-1-hd` (the gpt-4o-mini-tts case is split out into a v2 issue) - `src/agents/tts_google_agent.ts` - `src/agents/tts_kotodama_agent.ts` ## Return shape Each agent returns: ```ts return { buffer, usage: { provider: "openai" | "google" | "kotodama", model: usedModel, // openai: "tts-1" | "tts-1-hd"; google: voice tier; kotodama: voice/decoration ID inputChars: text.length, // Unicode code points; spaces + newlines included cached: false, }, }; ``` ## Notes - **Google TTS** charges spaces, newlines, and most SSML tags (except `<mark>`). mulmocast currently doesn't emit SSML, so `text.length` is correct. If SSML emission lands later, count includes tags too — no agent change needed. - **Kotodama** is subscription-billed externally; we just surface what was sent so the API/billing layer can do its own accounting. - **OpenAI tts-1 / tts-1-hd** are character-billed (see #1420 research). `gpt-4o-mini-tts` is token-billed and is handled by a separate issue. ## Acceptance - [ ] All 3 agents extended (backward compatible — existing consumers only read `buffer`) - [ ] Callback in `src/actions/audio.ts` pushes `usage` into `context.usageCollector` - [ ] Unit / mocked test asserting an `UsageRecord` with the correct `inputChars` is recorded - [ ] Cache hits not recorded (verify) ## Dependencies - #1416 (UsageCollector foundation) - Research closed via #1420