[chenghao-mou]

Add capability-gated STT flushing for manual commits and EOT. Only negative predictions delay flushing to half max_delay. Keep final-only EOT input, existing transcript triggers, and the audio-only stt_node contract.

ElevenLabs changes and tests are in #7287; other providers are deferred to #7989. Buffering custom nodes must disable manual_flush. Full-suite validation is blocked by server/credential fixtures.

Related: @IngLP's report #4087 and @longcw's capability proposal #4043.

Addresses AGT-3514 (Linear)

Initial prompt and agent context

Model: GPT-6

do we actually send _FlushSentinel to 11labs stt from VAD EOS?

okay, create a fix to flush STT with VAD EOS when STT supports it.

we can't change stt node contract, right? that means breaking change for any custom stt node implementation?

i guess it is okay, if they are buffering, they can disable by turn off manual flush.

so an ideal implementation is to flush in commit user turn (manual path) or to flush at EOT positive or some timeout when it is negative, does that sound right to you?

okay, let's do that, manual commit only uses flush if the STT supports that

we should keep it simple, just use the 1/2 value.

but we return 1.0 when there is an issue, right?

Keep final-only EOT input

can you move the 11labs changes to the other PR?