[Tony25032]

Fixes #603

Summary

Non-English dictation gets silently translated: a Chinese (or Korean, Japanese, …) transcript goes into refinement and an English rewrite comes back in transcript_refined. This implements the required fix proposed in #603 — inject the transcript language into the refinement prompt as a hard output constraint.

Root cause

The refinement system prompt (backend/services/refinement.py) is assembled from four English instruction blocks with English-only few-shot examples, and nothing anywhere pins the output language. With temperature=0.2, small refinement models (Qwen3 0.6B/1.7B/4B) pattern-match "the examples are English → the output is English" and treat the task as translate and clean. English dictation hides the bug completely, which is why it went unnoticed.

Reproduction (synthetic Chinese example; see also the Korean repro in #603)

• Raw STT (language=zh): 「嗯…那个…我们下周开会，啊不对周四周四下午两点开，然后嗯把报告发我一下」
• Refined before this patch: comes back as an English translation of the same content (observed across Qwen3 0.6B–4B), e.g. "Um… we have a meeting next week, no wait, Thursday at 2 pm, and um, send me the report." — the speaker's wording is silently lost.
• Refined after this patch: 「我们周四下午两点开会，然后把报告发我一下。」 — stays Chinese, fillers and the mid-utterance self-correction still resolved.

Changes

• refine_transcript() accepts the Whisper language code and forwards it to build_refinement_prompt().
• build_refinement_prompt() injects an output-language constraint: a named hard constraint when the language is known ("Your entire output MUST be written in Chinese … Never translate the transcript"), or a same-language rule when the code is unknown. This anchors the model against the English-only examples instead of relying on the model's language-mirroring instincts.
• refine_capture() forwards the capture row's language.
• New backend/tests/test_refinement_language.py pins the prompt-level guardrails (pure prompt-builder tests, no model required).

Scope

Deliberately limited to the language-preservation fix requested in #603. The issue's larger asks — localized few-shot examples/filler lists, per-language presets, and a prompt editor UI — are natural follow-ups; happy to work on them if this approach lands.

Testing

• pytest backend/tests/test_refinement_language.py (new)
• Existing suites are unaffected: build_refinement_prompt() keeps its previous behaviour when called without language, and refine_transcript()'s new parameter is optional with default None.

Happy to reword the injected constraint if maintainers prefer different phrasing.