perf: stream cloud media buffers

[JustYannicc]

What changed

• Streams ElevenLabs and Mistral Voxtral STT multipart uploads by writing the existing multipart parts sequentially instead of concatenating a full request body.
• Streams successful ElevenLabs TTS MP3 responses directly to the target temp file while keeping HTTP error text bounded.
• Adds focused cloud media buffering tests with before/after byte diagnostics for STT multipart copies and TTS response buffering.

Why

• Wave4 closes the remaining cloud STT/TTS media buffering gaps left after the optimized OpenAI Whisper upload path.
• Long Read/TTS sessions and large STT captures should avoid retaining avoidable full-body copies in the Electron main process.

Compatibility impact

• Preserves provider endpoints, request field order, request headers, multipart Content-Length accuracy, API key handling, Mistral timeout behavior, JSON/text parsing branches, and successful output semantics.
• ElevenLabs TTS still rejects empty audio and now removes partial streamed files on stream/empty-audio failure.
• Non-2xx ElevenLabs TTS/STT detected_unusual_activity messaging is preserved; error messages continue to expose the first 500 response characters.

How tested

• node --test scripts/test-cloud-media-buffering.mjs
• node --test scripts/test-ai-provider-whisper-upload.mjs
• node --test scripts/test-whisper-file-buffer-read.mjs
• ./node_modules/.bin/tsc -p tsconfig.main.json --noEmit --pretty false
• Codex LSP diagnostics for src/main/main.ts: no diagnostics found.

Performance evidence

• ElevenLabs STT test: legacy full multipart body copy would allocate 6,291,868 bytes; streamed path writes 6 buffers, reuses the 6,291,456 byte audio Buffer, and allocates 412 framing bytes.
• Mistral Voxtral STT test: legacy full multipart body copy would allocate 7,340,475 bytes; streamed path writes 6 buffers, reuses the 7,340,032 byte audio Buffer, and allocates 443 framing bytes.
• ElevenLabs TTS test: legacy response handling would retain a 5,767,168 byte MP3 Buffer before fs.writeFile; streamed path writes 5,767,168 bytes directly to a 5,767,168 byte temp file without a provider-frame Buffer.concat.

Stack validation

• Started from codex/perf-wave3-integration-stack-20260704 at commit 206eab8280a2eb91a57e4e74ea6df99e45a55cb4.
• Implementation branch: codex/perf-wave4-cloud-media-buffering-20260704.
• No coordinator markdown, playbooks, prompts, or generated locks are included.

Replaces

• Builds on and replaces the wave3 integration stack for the cloud media buffering portion of the perf work.

[JustYannicc]

Closing in favor of #601, which fully replaces this granular PR in the wave4 runtime/main consolidation. Literal validation now passes with /opt/homebrew/opt/node@22/bin/npm run build:main, and #601 is ready for review.