Fix English text incorrectly showing 'translated by omi' badge

[beastoin]

Summary

• Add should_persist_translation() guard that skips creating Translation objects when translated text is identical to source text (same-language no-op)
• Wire guard into _translate_segment() after API response — checks before persisting, prunes pending entry on skip
• Normalize locale tags consistently in update_from_translate_response() for robust comparison

Root cause

PR #5384 (translate cost optimization) removed the old if translated_text == segment_text: return guard. Combined with langdetect misdetecting short English text (e.g. "Transcription service." → fr), short segments bypass the lang cache and reach the API, which returns identical text that gets persisted with a translation badge.

Fix mechanism

should_persist_translation(source, translated, detected_lang, target_lang):

1. If normalized texts differ → persist (real translation happened)
2. If texts are identical AND detected_lang base matches target → skip (confirmed no-op)
3. If texts are identical AND no detection → skip (conservative: unchanged text = no-op)

Guard placement: after lang cache update (so cache still learns from API) but before Translation object creation.

Testing

• 86/86 unit tests pass (18 new tests)
• Level 1 live test: real Google Translate API — all pass (short foreign words translate correctly, short English no-ops skipped)

Files changed

• backend/utils/translation_cache.py — should_persist_translation() + _normalize_base_language() helper
• backend/routers/transcribe.py — import + guard in _translate_segment()
• backend/tests/unit/test_translation_optimization.py — 18 new tests

Deployment

• Type: Code-only change (no env vars, no config changes)
• Action needed: Deploy Backend to Cloud Run (gh workflow run gcp_backend.yml -f environment=prod -f branch=main)
• Cloud Run deploy rebuilds the image AND restarts backend-listen pods — no separate Helm upgrade needed
• No pusher/diarizer/VAD deploy needed — change is only in transcribe.py and translation_cache.py

Closes #5582

by AI for @beastoin

[greptile-apps]

Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.

[beastoin]

Level 1 Live Test: Local dev backend + real Google Translate API

Ran TranslationService.translate_text_by_sentence() against real Google Translate API with dev credentials. Verified full pipeline: langdetect pre-filter → API call → should_persist_translation guard.

Q1: Short foreign words (target=en) — all translate correctly

Source	API Result	Detected	Persist	Status
"Bonjour."	"Good morning."	fr	True	PASS
"Hola."	"Hello."	es	True	PASS
"Danke."	"Thanks."	de	True	PASS
"Merci."	"THANKS."	fr	True	PASS

Q2: Multiple short foreign sentences — all translate correctly

Source	API Result	Detected	Persist	Status
"Bonjour. Comment allez-vous?"	"Good morning. How are you doing?"	fr	True	PASS
"Hola. Gracias. Adiós."	"Hello. Thank you. Bye bye."	es	True	PASS
"Oui. Non. Merci beaucoup."	"Yes. Non. Thank you so much."	fr	True	PASS

Bug fix: Short English no-op correctly skipped

Source	API Result	Detected	Persist	Status
"Transcription service."	"Transcription service."	en	False	PASS
"Hey."	"Hey."	en	False	PASS
"123. 123."	"123. 123."	en	False	PASS
"Hello? Can you hear?"	"Hello? Can you hear?"	en	False	PASS

Control: Long English pre-filtered

Source	Pre-filter	Persist	Status
"Two three four five. The weather is nice today."	skip=True	False	PASS

Result: ALL PASS (86/86 unit tests + level 1 live test)

by AI for @beastoin

[beastoin]

lgtm

[beastoin]

Prod Verification (T+0)

Deploy confirmed: Cloud Run revision backend-00839-z94, backend-listen pods rolling out.

Translation pipeline: healthy

• Active sessions with translation enabled processing normally
• lang_cache_skips working (e.g. total=66, lang_skip=65, buffered=1 — English segments correctly skipped by pre-filter)
• Real translations flowing (e.g. total=88, buffered=84, translated=84 — foreign segments correctly translated)
• Translation API calls active with cache hits
• Zero translation errors in last 15min post-deploy

Sample translate_summary logs (prod)

total=66 buffered=1 translated=1 lang_skip=65  (mostly English, 1 foreign segment translated)
total=23 buffered=6 translated=6 lang_skip=17  (mixed session, all foreign segments translated)
total=88 buffered=84 translated=84 lang_skip=4  (heavily foreign session, all translated)
total=3 buffered=0 translated=0 lang_skip=3    (all English, correctly skipped)

Result: T+0 clean. Translation pipeline operating normally post-deploy.

by AI for @beastoin