Characterize callback-suppressed activation reason labeling.

Characterize read-error reason labeling.

Test capture_speech returns None on timeout without speech.

Characterize timeout-after-activation reason labeling.

Characterize timeout-without-activation reason labeling.

Timeout should be based on elapsed wall-clock time, not nominal timeout_chunks.

After text is yielded, ConnectionClosed should propagate (no replay retry).

Concurrent speaker feed + mic processing shouldn't crash.

Parakeet fills placeholder confidence values.

Multiple cumulative deltas should yield only the new portions.

After tool use, cumulative text resets for the next LLM step.

When wake_word.enabled is false, wake_word_model should be None after init.

DISABLE_SLEEP=1 should override config.

When mic = speaker (pure echo), output energy should decrease.

Verify the cancellation is not just marginal — should reduce energy significantly (at least 50% = 3dB) in a pure echo scenario after

Synthetic echo test: feed identical sine wave as both speaker and mic. After warmup, APM should reduce echo energy.

Final with no text should not yield a tail delta.

Empty model name should raise RuntimeError.

When wake_word.enabled is true, model should load.

Encoded image should not exceed max_size in either dimension.

Parakeet on GPU should transcribe 2s audio in under 500ms.

24kHz WAV should be resampled to 16kHz.

Explicit .onnx file path should load.

chat_final with text beyond last delta yields a tail CHAT_DELTA.

If final text == last delta text, no extra tail delta.

Flush is a no-op through ResamplingPyAudioStream (no get_read_available). ResamplingPyAudioStream intentionally does NOT expose get_read_avai

Test get_date returns current date.

Test get_time returns current time.

Test get_tool_handler retrieves existing tools.

Test get_tool_handler returns None for non-existent tools.

Calling _ensure_cuda_ld_path twice should not corrupt LD_LIBRARY_PATH.

chat.inject returns False on error response.

chat.inject must skip unrelated frames (health ticks, etc.) and find its ACK.

chat.inject returns True on ok response.

chat.inject omits label when not provided.

chat.inject must not accept a response for a different request id.

Sleep should use small chunks for responsive stopping.

Nonexistent model name should raise RuntimeError.

lifecycle:end is informational — must NOT terminate the stream.

If close happens before any text, retry should recover without duplicate lifecycle.