Patch base-image CVEs and unblock tts-qwen-cuda build

The v1.1.0 docker-publish workflow flagged 22 CVEs in the Alpine and
Ubuntu bases of the published images (2 critical OpenSSL + 20 high
across libpng, libexpat, musl, nghttp2, zlib, OpenSSL libssl3/
libcrypto3). All in OS-level packages; real exploit surface is
limited for our deployment because TLS terminates at CF Pages, not in
the container, but better to ship clean bases.

Add base-image security upgrades on top of each upstream:

  client/Dockerfile               apk upgrade --no-cache (Alpine)
  docker/tts-kokoro/Dockerfile    apt-get upgrade -y (Debian 13)
  docker/stt-whisper/Dockerfile   apt-get upgrade -y (Ubuntu)
  docker/tts-qwen-cuda/Dockerfile apt-get upgrade -y (Ubuntu 22.04)

Also unblock the tts-qwen-cuda build, which failed entirely on
v1.1.0. Two compounding issues:

  1. `sox` system binary missing. The Python `sox` package
     (transitive dep of faster-qwen3-tts) calls `sox --version`
     during setup.py. Add `sox` to the apt-get install line.

  2. Numpy bootstrap. The Python `sox` package's setup.py imports
     its own module to read __version__, which transitively imports
     numpy. Without numpy already on disk, pip's metadata-generation
     step for sox fails before sox itself can be installed. Split
     the pip install into three steps so numpy is present when
     sox's egg_info runs:

       pip install torch                  (existing pin)
       pip install numpy                  (NEW pre-seed)
       pip install faster-qwen3-tts ...   (the rest)

Verified by rebuilding tts-qwen-cuda locally on Apple Silicon (arm64
emulation) with --no-cache. The other three images use well-known
base-image upgrade patterns and rely on the existing CI workflow for
validation.

Author: michael-chipmates

client/Dockerfile

@@ -55,7 +55,10 @@ COPY docker-entrypoint.sh /docker-entrypoint.d/00-agora-config.sh
 # scripts run as the container's runtime user (not root), so the entrypoint
 # needs write access to the static-root directory to rewrite config.js.
 USER root
-RUN chmod +x /docker-entrypoint.d/00-agora-config.sh && \
+# Pull latest Alpine security patches on top of the upstream base
+# (openssl libcrypto3/libssl3, libpng, libexpat, musl, nghttp2, zlib).
+RUN apk upgrade --no-cache && \
+    chmod +x /docker-entrypoint.d/00-agora-config.sh && \
     chown 101:101 /usr/share/nginx/html /usr/share/nginx/html/config.js
 USER 101
 

docker/stt-whisper/Dockerfile

@@ -29,6 +29,8 @@ ENV ALLOW_ORIGINS=[\"*\"]
 # request doesn't 404. Skipped on subsequent starts (snapshot_download is
 # idempotent against the HF cache).
 USER root
+# Pull latest Ubuntu security patches on top of the upstream Speaches base.
+RUN apt-get update && apt-get upgrade -y && rm -rf /var/lib/apt/lists/*
 COPY entrypoint.sh /usr/local/bin/agora-init.sh
 RUN chmod +x /usr/local/bin/agora-init.sh
 USER ubuntu

docker/tts-kokoro/Dockerfile

@@ -23,6 +23,10 @@
 
 FROM ghcr.io/remsky/kokoro-fastapi-cpu:latest
 
+# Pull latest Debian security patches on top of the upstream base.
+USER root
+RUN apt-get update && apt-get upgrade -y && rm -rf /var/lib/apt/lists/*
+
 # Patch logging to WARNING level. Kokoro by default logs incoming request
 # text at DEBUG.
 RUN sed -i 's/"level": "DEBUG"/"level": "WARNING"/' /app/api/src/main.py && \

docker/tts-qwen-cuda/Dockerfile

@@ -14,8 +14,11 @@
 FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
-RUN apt-get update && apt-get install -y --no-install-recommends \
-      python3.11 python3.11-venv python3-pip ffmpeg curl ca-certificates \
+# `sox` is required at install time by faster-qwen3-tts (transitive dep
+# `sox` Python package calls the binary during setup.py). `apt upgrade`
+# pulls latest security patches for the Ubuntu 22.04 base.
+RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-recommends \
+      python3.11 python3.11-venv python3-pip ffmpeg sox curl ca-certificates \
     && rm -rf /var/lib/apt/lists/*
 
 RUN python3.11 -m venv /opt/venv
@@ -25,6 +28,13 @@ ENV PATH="/opt/venv/bin:$PATH"
 # faster-qwen3-tts. The README is explicit: torch<=2.5.0 has graph-capture
 # bugs in this path, and Blackwell (RTX 50xx) needs cu128 / 2.7+.
 RUN pip install --no-cache-dir torch==2.5.1 --index-url https://download.pytorch.org/whl/cu124 \
+    # Pre-seed numpy + sox before faster-qwen3-tts pulls them in. The Python
+    # `sox` package's setup.py imports its own module to read __version__,
+    # which transitively imports numpy. Without numpy already installed
+    # pip's metadata-generation step for sox fails before sox itself can
+    # be installed. Splitting the install ensures numpy is on disk when
+    # sox's egg_info runs.
+    && pip install --no-cache-dir numpy \
     && pip install --no-cache-dir \
          faster-qwen3-tts \
          fastapi==0.115.* \