STT diarization request silently degrades to flat transcription

[martsokha]

Context

`crates/nvisy-engine/src/extraction/stt/mod.rs:69`:

```rust
if diarization {
tracing::warn!(target: TARGET, "diarization not yet supported, skipping");
}
```

Caller asks for speaker diarization via `extraction.audial.diarization = true`, gets a flat single-block transcript instead — with no error, just a buried tracing::warn. Operator has no way to know their config silently degraded.

Fix shape

Three paths, easiest to hardest:

1. Fail at config-load time when `diarization = true` is set. One `Result::Err` with an explanatory message in `Extractors::from_config` or earlier. Operators see the failure immediately instead of after a run completes.
2. Surface in run output. Add a degradation field to the run's audit/metadata so downstream consumers can see "requested diarization, got flat transcript".
3. Wire diarization properly. Whisper's `verbose_json` + `timestamp_granularities="segment"` + a downstream speaker-clustering pass. Real feature work; siblings issue STT extraction stamps TimeSpan(0,0) on speech blocks #238 (TimeSpan(0,0)) and the TODO at `nvisy-agent/src/audio/stt/mod.rs:60`.

(1) is the immediate unblocker. (3) is the actual feature.

Related

• STT extraction stamps TimeSpan(0,0) on speech blocks #238 — same STT extractor, same upstream gap
• Punch list batch 6.5 in `.ignore/engine-polish.md`