QVAC-18524 fix: avoid Node-only Buffer in RN duplex RPC path

[Victor-Rodzko]

Note: be concise and prefer bullet points.

🎯 What problem does this PR solve?

• On RN/Hermes (iOS + Android) every transcribeStream() call from @qvac/sdk throws ReferenceError: Property 'Buffer' doesn't exist on the very first write.
• The new transcribe-stream-events-* e2e tests (added in QVAC-17282 feat[api]: expose whisper VAD and end-of-turn events in transcribeStream #1848 / QVAC-17282) cannot run on mobile at all today.
• Root cause: the bare-rpc duplex stream polyfill internally does Buffer.from(chunk, encoding) for string writes, but Hermes has no global Buffer — so requestStream.write(payload, "utf-8") blows up before any audio data reaches the worker.

📝 How does it solve it?

• Pre-encode the JSON request payload to a Uint8Array via TextEncoder in createDuplexSession so the polyfill's binary branch is taken on every runtime.
• Drop Buffer.from(string) from the profiled duplex response generator and yield strings directly — downstream chunk.toString() consumers work for both Buffer and string.
• Widen the public TranscribeStream*Session.write(audioChunk) signatures from Buffer to Buffer | Uint8Array so RN callers can pass binary slices directly instead of polyfilling Buffer.
• Stop wrapping Uint8Array slices in Buffer.from(...) inside the shared transcribe-stream test runner.
• Skip transcribe-stream-events-* on iOS under the existing TODO(QVAC-18460) since the underlying native Silero/Whisper crash is identical to the one already blocking transcription-* on iOS — flipping both skips will be a single follow-up commit when 18460 lands.
• Add a one-line rule to tests-qvac.mdc: no Node-only globals in tests/shared/ or tests/mobile/.

Type Changes

Additive type widening — all existing Buffer-based callers compile and behave identically.

// Before
interface TranscribeStreamSession {
  write(audioChunk: Buffer): void;
  // ...
}

// After
interface TranscribeStreamSession {
  write(audioChunk: Buffer | Uint8Array): void;
  // ...
}

Also widened on TranscribeStreamMetadataSession and TranscribeStreamConversationSession, and on the DuplexWritable / DuplexReadable internal RPC types.

Example RN/Bare-friendly usage:

import { transcribeStream } from "@qvac/sdk";

const session = await transcribeStream({ /* ... */ });
const pcm = new Uint8Array(audioBuffer); // straight from any binary source
session.write(pcm);                       // ✅ no Buffer needed
session.end();
for await (const text of session) console.log(text);

🧪 How was it tested?

• Local bun run lint + tsc --noEmit on packages/sdk — green.
• Local iOS run via npx qvac-test run:local:ios --filter transcribe-stream-events:
  • Before: ReferenceError: Property 'Buffer' doesn't exist ~70 ms after test start.
  • After: tests skipped on iOS with the QVAC-18460 reason; the JS-side wire-up no longer throws when run against a non-crashing native (verified locally before the iOS-specific Silero crash kicks in).
• Desktop transcribe-stream-events-* continue to pass — Buffer callers compile and execute unchanged thanks to the union widening.

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*

[lauripiisang]

@Victor-Rodzko before merge I'd ask to consider removing [api] tag and the "API Changes" from description though -- I'm not sure this constitutes a meaningful API change and will just pollute the changleog
Already did it