docs(nvidia-stt): correct the finalize rationale — Riva does endpoint itself

The comment claimed "Riva's own endpointing never fires". Live logs disprove
it: a final arrives mid-meeting with no Answer Now press, from Riva's own
endpointing.

What Riva will not do is release the utterance still IN FLIGHT when the user
asks for an answer — and that trailing fragment is usually the question. One
measured press: a 14-char final had already landed on its own; the 39-char
remainder, which was the actual question, only arrived once finalize closed
the call.

So the rotation is still required and the code is unchanged. Only the stated
reason was wrong, and wrong in a way that would mislead the next person into
thinking a no-op merely delayed finals rather than dropping the important one.

Author: evinjohnn

electron/audio/NvidiaNimStreamingSTT.ts

@@ -104,11 +104,17 @@ export class NvidiaNimStreamingSTT extends EventEmitter {
    *
    * The renderer calls this on every "Answer Now" to mean "transcribe what I
    * just said". Riva has NO flush control — StreamingRecognize is one long
-   * call, and half-closing it is the only way to make the server emit the
-   * final it is holding. That is why the original `stream.end()` produced a
-   * transcript, and why replacing it with a no-op produced NONE at all: the
-   * mic's silence suppressor sends keepalive frames rather than real silence,
-   * so Riva's own endpointing never fires and nothing else forces the final.
+   * call, and half-closing it is the only way to make the server release a
+   * final it is still holding.
+   *
+   * Riva DOES endpoint on its own for completed utterances; live logs show
+   * finals arriving mid-meeting with no press. What it will not do is release
+   * the utterance still in flight at the moment the user asks for an answer,
+   * and that trailing fragment is usually the question itself. Measured on one
+   * press: a 14-char final had already landed from normal endpointing, and the
+   * 39-char remainder — the actual question — only arrived once this closed the
+   * call. Replacing end() with a no-op therefore did not merely delay finals,
+   * it lost the one that mattered.
    *
    * So: ROTATE rather than close. End the current call (the server flushes its
    * final, which still reaches the listener — the 'data' handler emits