[android] Validate on-device ASR runtime bench (#120) on Polish audio

[grzanka]

[android] Validate on-device ASR runtime bench (#120) on Polish audio

Background

Issue #120's Android on-device ASR runtime spike (findings: docs/android-asr-runtime-bench.md,
PR #126) measured all four candidates — whisper.cpp, sherpa-onnx, Vosk, wav2vec2 — on a real
Pixel 7a, but every run so far is English-only. No Polish eval audio exists yet (blocked on
issues #63/#79/#30), so whisper.cpp's 89% E2E and sherpa-onnx's 87% E2E on-device numbers are not
yet confirmed to hold for Polish queries, even though both are advertised as multilingual models.

Why this matters

This project's stated goal is a bilingual (PL+EN) physicist assistant. The whole android runtime
decision in #120 was made on English-only evidence. If Polish accuracy is meaningfully worse
on-device — a real risk, since Whisper's per-language accuracy is known to vary — the "whisper.cpp
is the clear winner" conclusion in docs/android-asr-runtime-bench.md §9 could be premature for
half the target user base.

Scope

Once Polish eval audio exists (tracked by #63/#79/#30):

• Re-run the same on-device benchmark harnesses already built in bench/android/{whispercpp,sherpa-onnx}
  (Vosk and wav2vec2 don't need re-testing — both were already disqualified on English for
  reasons — closed vocabulary gap, no digit tokens — that would only be worse for Polish) against
  the Polish clips.
• Compare against whatever Polish desktop CPU baseline exists at that point.
• Land findings as a new section in docs/android-asr-runtime-bench.md (or a follow-up doc if
  it's a large enough addition), per this repo's spike-findings convention — not just a closing
  comment.

Non-goals

• No new runtime candidates — this validates the [android] Spike — measure on-device ASR runtime candidates (whisper.cpp, sherpa-onnx, Vosk) for latency/battery before committing to a native rewrite #120 winner(s) on a second language, it doesn't
  re-open runtime selection.

Blocked by

• [i18n] Polish-language support: voice dictation, deterministic NLU, and a UI language switch #63, [i18n] Define sentence/speaker counts for the Polish eval set (recorded vs. generated) #79, [eval] TTS-synthesized eval audio — scale ASR evaluation without recording sessions #30 (Polish eval audio)