…cascade

`assembly agent-framework` holds the same kind of live voice conversation as
`assembly agent`, but instead of talking to AssemblyAI's Voice Agent endpoint it
wires the three primitives together itself — Streaming STT -> the LLM Gateway ->
streaming TTS — exactly like the `agent-framework` init template does
server-side. Sandbox-only (streaming TTS has no production host).

New feature slice `aai_cli/agent_framework/`:
- engine.py: the cascade orchestrator (greeting, per-sentence TTS, barge-in,
  sliding-window history). The three network legs are injected through
  `CascadeDeps` (the `tts/session.py` seam), so it's unit-tested against fakes
  with no sockets, mic, or speaker.
- config.py / voices.py / text.py: per-run config + the streaming-TTS voice
  catalog presentation + sentence-splitting/history-trimming helpers.

New command package `aai_cli/commands/agent_framework/` follows the
options/run-split convention and reuses DuplexAudio/AgentRenderer,
client.stream_audio, llm.complete, and tts.session.synthesize.

Also clears pre-existing CodeQL `py/ineffectual-statement` findings in the
agent-framework / audio-transcription init templates and their tests (await-of
-name statements and a `...` Protocol stub) so the full local gate is green.

CI runs pytest with FORCE_COLOR, so the Rich usage-error panel carries ANSI
codes and width-based ellipsis — the long "Invalid value for
'--system-prompt-file'" substring no longer matched on the Windows/Linux
runners. Assert the behavioral distinction instead: a nonexistent
--system-prompt-file makes Typer reject before the body, so the sandbox guard
(the other exit-2 path) never runs and "sandbox" stays out of the output. This
still kills the exists=True mutant and is immune to ANSI/width.