fix(photon): preserve outbound voice attachment identity

[mooserini]

Summary

• Ship the outbound voice preparation needed by Photon/Spectrum 8.
• Materialize every outbound voice reply as a unique AAC/M4A file with matching audio/mp4 metadata and filename.
• Add source/upload identity logging and cleanup around the sidecar send path.
• Add real sidecar/Spectrum-boundary regression coverage so inbound CAF bytes cannot be reused as the outbound TTS attachment.

Related: #91129

Root cause

Hermes generated a distinct TTS MP3 and correctly extracted that MEDIA: path, but Spectrum 8's voice() path converted the bytes to M4A while preserving the source tts_*.mp3 upload identity. Photon/iMessage could then resolve the native audio message back onto the conversation's original inbound CAF voice memo.

This change creates a fresh voice-<UUID>.m4a object before calling voice(), with matching bytes, filename, and MIME metadata.

Verification

• node --check plugins/platforms/photon/sidecar/index.mjs plugins/platforms/photon/sidecar/voice-send.mjs
• pytest tests/test_photon_voice_dependencies.py tests/plugins/platforms/photon/test_sidecar_voice_runtime.py tests/plugins/platforms/photon/test_outbound_media.py tests/gateway/test_tts_media_routing.py -q — 19 passed
• pytest tests/plugins/platforms/photon -q — 134 passed
• npm ls --depth=0 — spectrum-ts@8.0.0, ffmpeg-static@5.3.0
• Manual Photon validation after gateway restart confirmed that the generated reply was received as Ara's voice rather than the inbound voice note.

[alt-glitch]

This was generated by AI during triage.

Related: #91129 and #65322 cover the complementary CAF recognition/dependency and outbound conversion/threading paths. This PR specifically separates the generated outbound voice attachment identity after conversion.