fix(stt): stop Whisper handlers finalizing turns on progressive audio

[Mayankpratapsingh022]

Fixes #412

Summary

• branch on vad_audio.mode in the four Whisper-family STT handlers so progressive
  audio yields PartialTranscription rather than Transcription
• add tests/test_whisper_progressive_transcription.py, covering all four backends

The diagnosis and the shape of the fix are from @whitepepper355 in #412, including the
follow-up comment identifying mlx_audio_whisper_handler.py as a fourth affected file.

Problem

In --mode realtime the VAD emits a progressive chunk while the user is still speaking,
then a final chunk once speech ends. BaseSTTHandler closes a turn on the first
Transcription it observes:

def before_emit_output(self, output: STTOut) -> None:
    if isinstance(output, Transcription):
        self._mark_completed_final_revision(output)

should_process_input subsequently rejects every later chunk carrying that
(turn_id, turn_revision) as input-after-final.

ParakeetTDTSTTHandler avoids this by inspecting the mode at
parakeet_tdt_handler.py:244. The four Whisper-family handlers never read
vad_audio.mode and yielded Transcription unconditionally, so the progressive chunk
closed the turn and the remainder of the utterance was discarded:

• STT/whisper_stt_handler.py
• STT/faster_whisper_handler.py
• STT/lightning_whisper_mlx_handler.py
• STT/mlx_audio_whisper_handler.py

Parakeet is the default STT and covers roughly 25 languages. The defect is therefore
reached only after switching to a Whisper backend, which any language outside that set
requires.

Why the mode check alone is sufficient

ParakeetTDTSTTHandler gates its progressive path behind enable_live_transcription.
An equivalent flag is not required in these handlers. s2s_pipeline.py:500 sets
vad_kw.enable_realtime_transcription from --enable_live_transcription, and
vad_handler.py:623 gates the progressive yield on that value. Progressive chunks
therefore exist only when live transcription is enabled, and a handler that receives one
is already operating in that mode.

Verification

Each backend was exercised with real model weights, driven by the real VADHandler
rather than synthetic chunk boundaries. The VAD emitted ['progressive', 'final'] in
every case.

Reference transcription of the complete utterance:

I think the weather today is really quite nice and I would like to go outside.

Backend	Model	Delivered to the LLM before	After
whisper	openai/whisper-small	I think	matches reference
faster-whisper	small.en, int8	I think.	matches reference
whisper-mlx	small	I think	matches reference
mlx-audio-whisper	whisper-small-mlx	I think	matches reference

Before, on each backend:

[progressive] -> LLM Transcription 'I think'
dropping stale STT input-after-final for turn=turn_1 rev=0
[final] REJECTED by should_process_input

After:

[progressive]  draft PartialTranscription 'I think'
[final] -> LLM Transcription 'I think the weather today is really quite nice and I would like to go outside.'

The dropping stale line falls from one occurrence per turn to zero, which is the
signal #412 predicted.

On this branch: 923 passed, 1 skipped. ruff check, ruff format --check and
mypy src/ report no issues.

Test design

The regression test drives the BaseSTTHandler hooks in run() order and leaves the
input queue empty between chunks. Enqueuing the whole utterance up front instead trips
the separate progressive-before-final guard in should_process_input, which discards
the progressive chunks and allows the test to pass against unfixed code. Reverting
src/ while retaining the test produces 8 failures and 8 passes.

The optional backends are stubbed at import time so all four are covered on Linux CI,
where faster-whisper and lightning-whisper-mlx are not installed.

Out of scope

Exercising the transformers handler against real weights also surfaces an IndexError
at whisper_stt_handler.py:120, where pred_ids[0, 1] reads a text token rather than a
language tag under transformers 5.6.2. That behaviour is #375 and is already addressed
by #378, so it is left untouched here.

This change does not attempt to quantify the added latency of running STT twice per
utterance, which #412 raises. Measuring it requires the full LLM and TTS chain.

[andimarafioti]

Thank you! This is an important fix. I haven't used whisper in a while, which is why I missed it. Codex found a few issues, but I want to merge this quickly, so I'll merge this to a branch before targetting main. I'll keep your commits in the history!