Add --save-transcript and --save-dir flags for transcript/audio file output

[alexkroman]

Adds transcript and audio file saving to assembly stream, with two modes:

• --save-transcript PATH: Write finalized turns to a plain-text file, one line per turn
• --save-dir DIR --name "Title": Auto-name both transcript and WAV under DIR/YYYY-MM-DD/ with a timestamped, slugged stem

Key changes

New modules:

• aai_cli/streaming/naming.py: Filename assembly for --save-dir (date bucketing, slug generation, path validation)
• aai_cli/streaming/transcript.py: TranscriptWriter class that appends finalized turns to a UTF-8 text file, flushing per turn for robustness against Ctrl-C

Modified modules:

• aai_cli/commands/stream/_exec.py:
  • Added save_transcript, save_dir, name fields to StreamOptions
  • New _reject_save_with_show_code() validation (save flags incompatible with --show-code)
  • New _resolve_save_targets() to assemble final paths from flags, with mutual-exclusivity checks
  • Wired transcript writer into StreamSession initialization
• aai_cli/streaming/session.py:
  • Added _finalized_turn_line() helper to format turns for the transcript (matching text renderer's speaker prefixes)
  • Added save_transcript field to StreamSession
  • Integrated TranscriptWriter into turn recording and LLM follow mode
• aai_cli/commands/stream/__init__.py: Added CLI options for --save-transcript, --save-dir, --name with help text and examples

Test coverage:

• tests/test_stream_session.py: Refactored to focus on session lifecycle and transcript saving; removed system-audio tests (moved to new file)
• tests/test_stream_system_audio.py: New file with all --system-audio family tests (parallel sessions, worker errors, mic/system wiring)
• tests/test_stream_exec.py: Added tests for transcript writing, file validation, and --save-dir auto-naming with date bucketing
• tests/test_streaming_naming.py: Unit tests for slug generation, path resolution, and directory creation
• tests/test_streaming_transcript.py: Unit tests for TranscriptWriter and path validation

Implementation details

• Transcript lines are written only for finalized, non-empty turns; partials and empty turns are skipped
• Speaker labels (diarization) are prefixed exactly as the text renderer does, so saved transcripts match on-screen output
• --save-dir auto-names files with a sortable timestamp (YYYY-MM-DD-HHMMSS) and optional slug from --name, bucketed under DIR/YYYY-MM-DD/
• Mutual exclusivity: --save-dir cannot combine with explicit --save-audio/--save-transcript paths, and cannot be used with --system-audio (two parallel streams can't share one WAV)
• File validation happens before streaming starts (missing parent directories fail cleanly up front)
• Flushing per turn ensures a Ctrl-C still leaves a complete transcript of turns seen so far
• Test file split: system-audio tests moved to dedicated test_stream_system_audio.py to keep test_stream_session.py under the 500-line gate

https://claude.ai/code/session_01UcH9QQmiZXviKzCpof7gJ4