EchoScript : Fixes : Why Persian/Portuguese/Turkish were failing , Korean removed from Marian targets , Multiple clicks , UI improvements

EchoScript : 20260701 0455 - Bug 1 (Persian missing) - the real root cause: The HF Hub returns 401 Unauthorized (not 404) for nonexistent repos when requests are unauthenticated. The fix: Replaced all live Hub checks with KNOWN_MARIAN_PAIRS - a static, hand-verified table of every (source, target) pair and its exact repo suffix. No network calls, no rate limits, no 401/404 ambiguity. Bug 2 (loading spinner on cached languages) - the real cause: generate_translations was a regular function, so Gradio held the loading spinner on every output component in translation_stage_outputs for the entire duration of the slowest new translation. The fix: Converted to a Gradio generator that yields twice: once immediately (pass 1, all cached results shown instantly, new languages left as no-op gr.update()), then once more per new language as each one finishes (pass 2).

EchoScript : 20260630 0857 - _marian_model_exists() , _safe_marian_exists() , HF_TOKEN/HUGGINGFACE_HUB_TOKEN

Helsinki-NLP published the English→Japanese model as opus-mt-en-jap - using jap, not the standard code ja - while the reverse direction (opus-mt-ja-en) correctly uses ja. That's an inconsistency in their own catalog, not my logic, but my code assumed standard ISO codes everywhere, so it was 404ing on a model that actually exists.

EchoScript : 1 & 2 - the static 4-language list was the real bug. 3 & 4 - already correct, verified by audit.

What changed:

EchoScript : made it automatic, not exclusive: TranslationService picks anthropic when ANTHROPIC_API_KEY is set, otherwise falls back to marian (fully offline/free) - overridable via ECHOSCRIPT_TRANSLATION_BACKEND=anthropic|marian. So EchoScript still works with zero API key, but gets better translation quality and zero missing-pair errors the moment you set one.

EchoScript : services\translation.py , Fix: services/translation.py now loads MarianMT directly via AutoModelForSeq2SeqLM + AutoTokenizer and calls model.generate() itself, bypassing the removed pipeline() task wrapper entirely. This is the path the official MarianMT docs use now too, so it's not a workaround you'll need to revisit on the next transformers release.

EchoScript : New Architecture