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

image/spark_embed/Dockerfile

@@ -0,0 +1,36 @@
+# spark-embed — dense embeddings (bge-m3) + reranker (bge-reranker-v2-m3)
+# Built FROM the NGC PyTorch image that is already proven to run on the DGX
+# Spark's GB10 (sm_121) GPU — the same base behind our vLLM and Kokoro work.
+#
+# Why not HF Text Embeddings Inference (TEI)? As of 2026 TEI ships no arm64
+# CUDA image (all *-cuda tags are amd64-only), so it won't run on the Spark.
+# Building on NGC torch sidesteps that AND avoids torchaudio (the dependency
+# that sank the WhisperX attempt). bge-m3 + the reranker are XLM-RoBERTa
+# encoders — no flash-attn, no torchaudio, just SDPA attention on torch.
+FROM nvcr.io/nvidia/pytorch:25.11-py3
+
+WORKDIR /app
+
+# Hard-pin the NGC torch version in a constraints file so pip CANNOT replace it
+# while resolving sentence-transformers. NGC's torch carries a local version
+# string (…nv25.11) not on PyPI; pinning it makes pip treat the already-installed
+# build as satisfying the requirement instead of pulling a PyPI wheel that
+# wouldn't have sm_121 kernels. (Same technique as the v0.12.0 torch-ABI work.)
+# transformers is NOT preinstalled in this NGC base, so it installs fresh from
+# PyPI; we cap it (<5) so a future major can't silently change loading behavior.
+RUN python -c "import torch; \
+open('/tmp/constraints.txt','w').write('torch==%s\n' % torch.__version__)" \
+ && cat /tmp/constraints.txt \
+ && pip install --no-cache-dir -c /tmp/constraints.txt \
+      "sentence-transformers>=3.0" "transformers<5" "fastapi>=0.115" "uvicorn[standard]>=0.30"
+
+COPY main.py /app/main.py
+
+# Persist HuggingFace model downloads (bge-m3 ~2.3GB + reranker ~2.3GB) on a
+# mounted volume so container recreates don't re-download.
+ENV HF_HOME=/data/hf
+ENV DENSE_MODEL=BAAI/bge-m3
+ENV RERANK_MODEL=BAAI/bge-reranker-v2-m3
+
+EXPOSE 8088
+CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8088"]