v0.13.0:4 - redaction gateway, embeddings proxy, expanded audio API

- Add redaction gateway (redaction_gateway.py, redaction/ scrub + tests)
- Add embeddings proxy and spark_embed service (Dockerfile + main.py)
- Expand audio_proxy with speaker-aware handling; deep_health/health/server updates
- Package: configureSparks action + sparkConfig model updates, manifest/main wiring
- Docs: AUDIO_API, EMBEDDINGS, REDACTION_GATEWAY; HANDOFF and runbook/known-issues refresh

package/startos/versions/v0_1_0.ts

@@ -1,10 +1,10 @@
 import { VersionInfo, IMPOSSIBLE } from '@start9labs/start-sdk'
 
 export const v0_1_0 = VersionInfo.of({
-  version: '0.13.0:3',
+  version: '0.18.0:0',
   releaseNotes: {
     en_US:
-      'v0.13.0:3 — chat-completions proxy. Adds POST /v1/chat/completions (and /v1/completions for the legacy endpoint) to Spark Control that forwards to whichever vLLM is currently loaded on Spark 1. Supports SSE streaming when stream=true in the request body. Request body is passed through unchanged — any vLLM-supported field works (model, messages, max_tokens, temperature, response_format, chat_template_kwargs, tools, ...). Closes the last gap that forced clients to know about both Spark Control AND the direct vLLM URL — recap-relay and friends can now use one trusted host for everything (transcribe, diarize, analyze) with one cert and one allowlist. 30-min request timeout to accommodate large-context completions. No parakeet container changes; no Reapply patches needed.',
+      'v0.18.0 — dual-channel mode for POST /api/audio/label-merge. Instead of one mixed-mono file, a caller (Ten31 Transcripts) can send two sample-aligned tracks: mic_file (the local user) + system_file (everyone else, from screen capture). Rather than force the diarizer to re-disentangle a mono mix (which over-segments — proven: a stereo clip of 2 clean voices returned 3 speakers), we split the problem so each model gets the easiest mono input. The mic track yields the local user\'s words, gated to windows where the mic is genuinely the user speaking (mic louder than system — a self-VAD computed server-side per-window, or supplied via self_vad); this gate is load-bearing because the mic picks up the remote audio as quiet bleed. The system track is diarized (only has to separate the remote people) and named via the visual timeline + voiceprints. The user\'s clean voiceprint is enrolled from the mic track and injected into the voiceprint library, so a system cluster that is the user dialed in from a second device (dual-login) resolves to the user, not a stranger. Validated on a real misattributing call: fixes both mono-mix misattributions, recovers the dropped-to-Unknown local line, and correctly splits overlapping speech (two people saying "Hello" at once) that the coarse ground truth itself conflated. New form fields: mic_file + system_file (dual mode), self_name, self_vad (optional). The mono file path is unchanged and fully backward-compatible. Response gains a "mode" field (mono | dual_channel). Known limit: if loud remote bleed masks a quiet local word, the mic-track ASR may miss it — mitigated by a cleaner mic (headphones) or future echo-cancellation. See docs/AUDIO_API.md.',
   },
   migrations: {
     up: async ({ effects }) => {},