fix(transcription): shared cross-provider hallucination filtering (follow-up to #679)

[Hus-Mek]

Follow-up to #679, split out per review so the VAD fix and this transcript-output policy change are reviewed independently.

What this fixes

Hallucination filtering only ran on local Whisper, and only knew English.

clean_repetitive_text and is_meaningless_output were private associated functions on WhisperEngine. The Parakeet arm and the provider (cloud) arm in worker.rs both stored result.text.trim() verbatim, so they got no filtering at all. The denylist was also entirely English ("thank you for watching", "like and subscribe", "applause").

Net effect: an English meeting got cleaned and a non-English meeting did not. A real 26-minute Arabic recording stored these unfiltered:

stored text	what it is
اشتركوا في القناة	"subscribe to the channel"
ترجمة نانسي قنقر	a subtitler credit, on a 0.15 s clip
شكرا x14	standalone - the Arabic analogue of Whisper's English "Thank you." attractor

What changed

• new audio/transcription/text_cleanup.rs - the cleanup routines extracted into one module and called from all three engine arms (local Whisper, Parakeet, remote providers), with Arabic patterns added alongside the English ones and grouped by language so further languages are additive.
• whisper_engine.rs - the now-shared code removed; delegates to the new module. (This is why the diff is move-heavy: ~155 lines relocated out of here.)
• worker.rs - the Parakeet and provider arms now call text_cleanup::clean_transcript_text instead of storing raw text. No change to the confidence gate - the 0.3 threshold removed by Fix: stop dropping transcripts below confidence threshold (#675) #681 is not touched or reintroduced here.
• transcription/mod.rs - registers the new module.

Behaviour, and the #675 correction

An earlier version of this work (in the original #679) included a standalone-filler denylist that dropped "thank you", "okay", "thanks", "you", "bye" and Arabic شكرا when they were a whole segment. That was reverted before this split: nothing in the text distinguishes شكرا decoded from 300 ms of silence from شكرا a participant actually said, so the list traded a cosmetic problem for a semantic one - and #675 is that exact failure already reported. Short-clip hallucination belongs in the provider layer gating on Whisper's own avg_logprob / compression_ratio, not in a text denylist.

What this PR filters by text is only unambiguous multi-word boilerplate (اشتركوا في القناة, subtitler credits, thanks for watching, like and subscribe) - phrases no meeting contains, so there is no genuine reading to lose.

Degenerate repetition is still caught language-agnostically: the repetition ratio is measured on the original text rather than after collapsing (measuring after collapsing hid it - a collapsed run leaves one word, ratio zero). شكرا شكرا شكرا شكرا شكرا is still discarded without the code needing to know what the word means.

Verification

$ cargo test --lib text_cleanup

running 14 tests
test audio::transcription::text_cleanup::tests::drops_arabic_subscribe_boilerplate ... ok
test audio::transcription::text_cleanup::tests::drops_arabic_subtitler_credit ... ok
test audio::transcription::text_cleanup::tests::drops_english_boilerplate_as_before ... ok
test audio::transcription::text_cleanup::tests::empty_and_whitespace_are_empty ... ok
test audio::transcription::text_cleanup::tests::discards_fully_degenerate_repetition ... ok
test audio::transcription::text_cleanup::tests::keeps_arabic_thank_you_inside_a_real_sentence ... ok
test audio::transcription::text_cleanup::tests::collapses_repeated_phrases ... ok
test audio::transcription::text_cleanup::tests::collapses_repeated_words ... ok
test audio::transcription::text_cleanup::tests::keeps_code_switched_arabic_english ... ok
test audio::transcription::text_cleanup::tests::keeps_ordinary_arabic_speech ... ok
test audio::transcription::text_cleanup::tests::keeps_short_acknowledgements_issue_675 ... ok
test audio::transcription::text_cleanup::tests::keeps_short_genuine_replies ... ok
test audio::transcription::text_cleanup::tests::keeps_tatweel_stretched_words ... ok
test audio::transcription::text_cleanup::tests::single_character_runs_are_junk ... ok

test result: ok. 14 passed; 0 failed; 0 ignored; 0 measured; 212 filtered out

keeps_short_acknowledgements_issue_675 asserts that OK, Yes, No, Sure, Got it, Thanks, Okay., Bye plus شكرا, شكراً, شكرا لك, نعم, طيب, لا and a tatweel-stretched شكـــرا all survive cleaning unchanged - i.e. this PR does not make #675 worse from a second direction.

The 14 tests cover both directions deliberately: the artifacts that must be removed, and ordinary Arabic, code-switched Arabic/English, and short genuine replies that must survive.

Note on diff size

The diff is ~506 changed lines, but ~155 of the deletions are code relocated verbatim out of whisper_engine.rs into the new module; net new logic is small. Happy to split further if you'd prefer the extraction and the cross-provider wiring in separate commits, but they don't stand alone cleanly.