Fix duplicate queued Codex prompts

[Time4Mind]

Summary

• enforce at-most-once full prompt submission after Codex TUI acceptance
• retain only the safe Enter retry while the exact text remains in the input field
• stop treating delayed transcript consumption as delivery failure
• remove the redundant post-dispatch voice verifier and false alert path
• keep the per-session inbound FIFO moving while Codex is busy

Reproduction fixed

A second voice prompt queued while Codex was still working could remain absent from the rollout for longer than 15 seconds. The delivery guard retyped it, so Codex later consumed two identical prompts and Telegram reported a false delivery failure.

Validation

• 1046 pytest tests passed
• Ruff check and format check passed
• Pyright: 0 errors
• module-size budgets passed