fix: detect the language on the offset window without VAD

[rubas-agent]

Before

Without VAD, build_params hands the window to whisper.cpp:

if req.vad_model_path.is_none() {
    if let Some(o) = req.offset_ms {
        params.set_offset_ms(u32_to_i32(o));
    }
    if let Some(d) = req.duration_ms {
        params.set_duration_ms(u32_to_i32(d));
    }
}

The two new integration tests on the origin/main Rust (source-built NIF):

1) test offset_ms detects the language on the window, not on the start of the buffer
   language: "nn"            # expected "en"
2) test offset and duration window the timeline without VAD
   List.last(segs).end = 10.399999618530273   # limit 8.5

Problem

whisper.cpp v1.9.4 detects the language at offset 0 (whisper_full_with_state, whisper_lang_auto_detect_with_state(ctx, state, 0, ...)). It reads offset_ms only after detection. So transcribe/3 with offset_ms: 30_000 decoded the window in the language of the first 30 s of the buffer. A :duration_ms window also returned segments past its end. The VAD path and transcribe_slice/4 already detect on the window, so the three ways to ask for one window gave different results.

After

transcribe_one cuts the window from the samples in Rust on both paths. whisper.cpp sees only the window.

Check	origin/main	Branch
Language of a 30 s noise prefix plus jfk, offset_ms: 30_000	"nn"	"en"
Last segment end, offset_ms: 2_000, duration_ms: 6_000	10.4 s	at most 8.5 s
Integration suite, source-built NIF	2 failures	86 passed

Cause, design, tests

Cause. native/whisper_cpp_native/src/transcribe.rs:build_params passed set_offset_ms/set_duration_ms to whisper.cpp when VAD was off. whisper.cpp windows the decode with them, but not the language detection.

Design. The VAD path already cut the window from the samples before it ran. Now the non-VAD path does the same: samples.truncate(win_end); samples.drain(..win_start). One code path owns the window, so detection and decode cannot disagree.

• window moved from vad.rs to transcribe.rs, because VAD no longer owns it. The set_offset_ms/set_duration_ms block is gone.
• Without VAD, segment and word times shift by win_start / 16000, so they stay absolute on the whole buffer. The VAD ranges already carry the window origin.
• An empty window returns TranscriptionResult::empty before whisper runs. The mel reflect pad reads out of bounds on 0 samples.
• duration_s stays the duration of the whole buffer, as before.

Tests.

• test/integration_test.exs: "offset_ms detects the language on the window, not on the start of the buffer". On the multilingual tiny model, 30 s of seeded white noise before jfk. transcribe/3 with offset_ms: 30_000 must return "en" and the same text as transcribe_slice/4 on the same window. Digital silence does not show the bug, because tiny detects "en" on silence.
• test/integration_test.exs: "offset and duration window the timeline without VAD". offset_ms: 2_000, duration_ms: 6_000 gives first.start >= 2.0 and last.end <= 8.5.
• Cargo test window_selects_offset_and_duration_in_samples.
• Both integration tests fail on origin/main and pass on the branch. task check passes: fmt, compile with warnings as errors, credo, clippy -D warnings, 195 mix tests, cargo tests, zizmor.

Docs: the :offset_ms/:duration_ms entry in lib/whisper_cpp.ex, and an [Unreleased] Fixed entry in CHANGELOG.md.