fix(tts): persist stepfunVoiceId on Character + harden probe race

Two follow-ups from pr-agent review of #79:

1. director.ts voicePromises built a Character WITHOUT stepfunVoiceId, so
   on a StepFun server the client (which omits the voice payload to save
   FOT) echoed back only voiceDescription — and the server re-scored via
   pickStepfunVoiceId every beat instead of honoring the LLM pick. The
   whole "CharacterDesigner picks a preset id" mechanism was effectively
   bypassed on live StepFun sessions (it only worked for prebaked cards,
   which carry stepfunVoiceId in their JSON). Persist stepfunVoiceId onto
   the Character so the client→server round-trip keeps the LLM selection.

2. fetchBeatAudio's null-provider branch (probe pending) required
   speaker.voice and silently dropped a stepfun-only speaker. Accept any
   synthesizable source (voice | stepfunVoiceId | voiceDescription) so a
   slow getTtsProvider probe can't drop audio during the first scene's
   fetch window. The server resolveVoice normalizes regardless of which
   fields arrive.

lib/engine/director.ts

@@ -308,6 +308,9 @@ export async function directScene(
   // On the StepFun path, thread the LLM-selected stepfunVoiceId from the card
   // into provision — it lets stepfunProvision honor the catalog pick instead
   // of falling back to the keyword scorer (same network cost: still zero).
+  // ALSO persist it onto the Character so the client can echo it back on a
+  // StepFun server (where it skips the ~220KB voice payload) and the server
+  // resolveVoice honors the LLM pick at synth time instead of re-scoring.
   const voicePromises = cards.map((card) =>
     provisionCharacterVoice(config, card.voiceDescription, card.name, {
       stepfunVoiceId: card.stepfunVoiceId,
@@ -316,6 +319,7 @@ export async function directScene(
         name: card.name,
         voiceDescription: card.voiceDescription,
         voice,
+        stepfunVoiceId: card.stepfunVoiceId,
       }),
     ),
   );