fix(cli): preserve voice transcript intent

[huseonglee-spec]

Summary

• stop prepending a hidden length/style instruction to CLI voice transcripts
• preserve voice-input provenance for command handling without altering the text sent to the agent
• keep API-local model/skill annotations and their clean persistence override intact
• replace the obsolete voice-prefix persistence fixture with a generic API-local annotation fixture

Root cause

PR NousResearch#1299 (commit a78249230) introduced an API-local prefix that forced voice replies to be conversational, limited them to 2–3 sentences, and prohibited Markdown/code blocks. That changes user intent solely because input came through STT and can truncate technical answers. The per-message sentinel added for NousResearch#65827 remains in place, so typed voice-control commands are still distinguishable from genuine STT output.

Test plan

• python -m pytest tests/cli/test_cli_interrupt_ack_race.py tests/tools/test_voice_cli_integration.py tests/run_agent/test_run_agent.py -o 'addopts=' -q
• repository search confirms the removed instruction and _voice_prefix no longer exist
• git diff --check