generate() ran unlocked against a std::span pointing into audioBuffer_,
relying on the vector's reservation never being exceeded. Unbounded
streamInsert from JS could grow the buffer past capacity, trigger
reallocation, and invalidate the span. Take a local copy under the
lock instead so the inference operates on stable data.

Previously `lastMerged.end = current.end` would shrink the merged
segment if a non-monotonic input arrived (current.end < lastMerged.end).
postprocess() doesn't produce such input today, but the safer form
removes the hidden invariant.

isStreaming stayed true after the native stream() resolved (whether
normally or via error), so subsequent code relying on the flag saw
stale state. Reset it in a finally block alongside the wake/finished
bookkeeping.

Mutating the in-flight `options` to flip `useVAD` off before the
final `finish()` call worked but left a footgun for anything later
that reads back `options.useVAD`. Build a local copy with the
override instead.

The function only reads from the span. Tagging it const signals
intent and matches the equivalent OnlineASR::insertAudioChunk
signature on the STT side.

`||` coerced an explicit `0` to the default 500. Switch to `??`
so callers can pass 0 to disable the margin.

Explain what the 1.2 multiplier means — widens the VAD merge window
relative to the user-configured detectionMargin so brief
intra-utterance silences don't split a single utterance into
separate segments.

`OnlineASR::process` computes the silence-trim cut as a `size_t`
subtraction of these two constants. If either is tweaked such that
the ordering inverts, the subtraction wraps and the subsequent
`erase` reads past the buffer. Lock the invariant in at compile
time.

SpeechToText gained a 4th positional `vadSource` argument; pass an
empty string at all 9 existing call sites so the test still exercises
the no-VAD path. Add the new VAD sources to the CMake target so the
binary links.

mergeSegments: empty input, single-segment passthrough, distant
segments stay separate, close/adjacent segments merge, overlapping
shorter inner doesn't shrink the result, mixed sequence merges
only adjacent close pairs.

stream/streamInsert/streamStop: stream() loop exits promptly on
streamStop, streamInsert while streaming doesn't crash, concurrent
stream() throws StreamingInProgress, and stream can be restarted
after a stop.

Covers the new PR behavior:
- valid vadSource constructs without throwing
- invalid vadSource fails loudly
- one-shot transcribe() is unaffected when VAD is loaded
- stream(useVAD=true) on a model built without VAD throws
- stream(useVAD=true) over pure-silence audio drives the VAD branch
  of OnlineASR::process and exits cleanly via streamStop()

Also register fsmn-vad in run_tests.sh so the SpeechToTextTests
runner pushes the VAD model alongside the Whisper artifacts.

Required by the const-span signature of VoiceActivityDetection::
streamInsert. Without this, ModelHostObject's template instantiation
of synchronousHostFunction<&VoiceActivityDetection::streamInsert>
references an undefined symbol at link time on Android.

Mirrors the existing std::span<float> specialization; the underlying
getTypedArrayAsSpan<float>() helper returns a span over the same
storage, which converts implicitly to span<const float>.