fix hydrate list and add no VAD

[minh-hoque]

Summary

This PR fixes false NO_RESPONSE / EMPTY_RESPONSE failures in OpenAI Realtime rehydrated runs (especially under parallel load) and hardens the manual --disable-vad workflow.

Rehydrate (proper response.create(input=...))

• Manual OpenAI Realtime rehydration no longer depends on flattening prior turns into system-prompt text.
• We now construct a structured Realtime input list from the golden history:
  • message(user)
  • function_call / function_call_output (when present)
  • message(assistant)

Per target turn, response.create includes

• Rehydration prefix items
• The committed live user audio item reference
• Pending tool-result items (for continuation turns)

No-VAD manual flow (--disable-vad)

• Server VAD is disabled (turn_detection=false).
• The client manually drives turn completion:
  1. input_audio_buffer.commit
  2. wait for input_audio_buffer.committed
  3. send response.create(input=...)
• Tool continuation is deferred until response.done, then resumed via manual response.create(input=...).

Stability hardening

• Added in-flight/debounce guards so duplicate stop events cannot trigger overlapping response.create.
• conversation_already_has_active_response is now recoverable in manual mode (non-fatal), preventing session collapse.
• Extended the active-output watchdog to reduce false NO_RESPONSE / EMPTY_RESPONSE.

Judge Alignment Policy Update (Rehydrated Runs)

• Disabled cross-turn realignment for rehydrated runs.
• Runtime-based detection: runtime.json.mode == "rehydrated".
• Rehydrated judge outputs now explicitly include:
  • realignment_applied = false
  • mode-specific judge_version (rehydrated no-realignment variant)