Add assembly dictate: hotkey-driven push-to-talk transcription

[alexkroman]

Implements the assembly dictate command for real-time, hotkey-driven transcription using the AssemblyAI Sync STT API. This is a new interactive command that lets users press Enter to record audio and get instant transcripts back without polling.

Key Changes

New modules:

• aai_cli/sync_stt.py: HTTP boundary for the Sync STT API (sync.assemblyai.com). Handles multipart PCM + JSON config uploads, normalizes errors into the CLIError hierarchy, and parses transcript responses. Stays Rich-free per the import-linter contract.
• aai_cli/hotkey.py: Single-keypress input via termios cbreak mode. Scopes terminal state changes to a with block, handles POSIX-only constraints gracefully (raises clean CLIError on Windows), and supports both blocking reads and zero-timeout polls for responsive recording.
• aai_cli/dictate_exec.py: Session orchestration for the dictate command. Implements the push-to-talk loop: wait for hotkey → record mic → transcribe → repeat. Supports language/prompt/word-boost flags, duration caps, and both human-readable and JSON output modes.
• aai_cli/commands/dictate.py: Typer CLI surface. Maps argv flags to DictateOptions and delegates to run_dictate.

Test coverage:

• tests/test_sync_stt.py (222 lines): HTTP boundary tests covering request shape, multipart encoding, error normalization (auth, rate limits, audio validation), response parsing, and fallback error detail extraction.
• tests/test_hotkey.py (119 lines): Terminal I/O tests using real pty pairs, verifying cbreak scoping, single-keypress reads, EOF handling, and clean error modes for non-TTY/non-POSIX platforms.
• tests/test_dictate_exec.py (312 lines): Session behavior tests with fully injected boundaries (scripted keys, canned PCM, mocked HTTP). Covers hotkey toggles, recording stops, duration caps, language/prompt handling, JSON mode, and Ctrl-C cleanup.
• tests/test_dictate_command.py (84 lines): CLI surface tests verifying argv → options mapping and the terminal-validation-before-credentials ordering.

Integration:

• Updated aai_cli/environments.py to add sync_base URL per environment (production and sandbox).
• Registered dictate command in aai_cli/main.py under the "Run AssemblyAI" workflow group.
• Updated help snapshot tests and tests/_snapshot_surface.py to include the new command.
• Updated .importlinter to allow aai_cli.dictate_exec as a source module.

Notable Implementation Details

• Boundary injection for testability: The session is driven by constructing DictateOptions and injecting three boundaries (TerminalKeys, MicrophoneSource, sync_stt.transcribe_pcm), so no test needs a real terminal, microphone, or network.
• Responsive recording: In-recording key polls use zero timeout (non-blocking) so audio chunks never stall behind keyboard reads; idle waits use None (blocking).
• Validation ordering: Terminal validation (TerminalKeys entry) happens before credential resolution, so piped stdin surfaces as "needs a terminal" rather than triggering a login prompt.
• Error normalization: All HTTP failures (auth, rate limits, audio validation) are normalized into the CLIError hierarchy with appropriate exit codes and suggestions.
• Immutable options: DictateOptions is a frozen dataclass, preventing accidental mutation during the session.
• Language handling: Comma-separated language codes are parsed into a list for code-switching audio; blank input is treated as unset.
• Prompt + language warning: When both --prompt and --language are set, a warning notes that the server ignores language_code with a custom prompt.

https://claude.ai/code/session_01FCXQLAyo8xpZiXrQ7hCMAf