build(docker): run as non-root and pin base images by digest

Dockerfile

@@ -1,12 +1,27 @@
-FROM python:3.12-slim
+# Base pinned by digest (the multi-arch index digest of python:3.12-slim, so amd64 and arm64 both
+# resolve); the tag is kept for readability only. Refresh both together:
+#   docker buildx imagetools inspect python:3.12-slim   → the top-level "Digest:" line
+FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1
 WORKDIR /app
 COPY pyproject.toml README.md LICENSE ./
 COPY passiveworkers/ passiveworkers/
 # [extract] = trafilatura, for real main-content/date extraction on page evidence. The bare
 # install degrades silently to a regex strip — this image is the README-recommended "just
 # works" path (research-desk AND coordinator both build from it), so it ships the better
-# extractor by default.
-RUN pip install --no-cache-dir -e '.[extract]'
+# extractor by default. A regular (non-editable) install: the package lands in site-packages,
+# owned by root and read-only to the runtime user below. (DL3013 wants a pinned version, but this
+# installs the local project, not a registry package.)
+# hadolint ignore=DL3013
+RUN pip install --no-cache-dir '.[extract]'
+# Run as the unprivileged user `pw`, by numeric ID so `runAsNonRoot`-style checks can verify it.
+# UID/GID 1000 matches docker-compose.yml's PW_UID/PW_GID default, so bind-mounted
+# ~/.passiveworkers files stay owned by the host user on Linux. /app/data is created here, owned by
+# that user, so a fresh named volume mounted there (the coordinator's DB) inherits a writable owner
+# instead of root.
+RUN groupadd --gid 1000 pw \
+    && useradd --uid 1000 --gid pw --create-home --shell /usr/sbin/nologin pw \
+    && install -d -o pw -g pw -m 0700 /app/data
+USER 1000:1000
 ENV PW_SERVE_HOST=0.0.0.0
 EXPOSE 8770
 CMD ["pworkers", "serve"]

docker-compose.yml

@@ -11,17 +11,20 @@
 # Docker Engine (20.10+) `host.docker.internal` now resolves via the `extra_hosts` entry below.
 # Running `pworkers` directly on the host with only SearXNG in Docker is still the simplest setup.
 #
-# Note (Linux only): the image has no USER, so it runs as root by default — harmless on Docker
-# Desktop (its VirtioFS/gRPC-FUSE layer remaps bind-mount ownership to your host user), but on
-# plain Linux Docker Engine a bind mount does NOT remap UIDs, so container root would write
-# root-owned (and for 0600 credential files, root-only-readable) files into your shared
-# ~/.passiveworkers, locking out your host-native `pworkers` CLI. PW_UID/PW_GID below default to 1000
-# (the common first-user UID on single-user Linux boxes) — override them with `id -u`/`id -g` if
-# yours differ, e.g.: PW_UID=$(id -u) PW_GID=$(id -g) docker compose up -d research-desk
+# Note (Linux only): the image runs as the unprivileged user `pw` (UID/GID 1000), never root. On
+# Docker Desktop that's moot (its VirtioFS/gRPC-FUSE layer remaps bind-mount ownership to your host
+# user), but on plain Linux Docker Engine a bind mount does NOT remap UIDs, so the container's UID is
+# what owns (and, for 0600 credential files, alone can read) the files it writes into your shared
+# ~/.passiveworkers. PW_UID/PW_GID below default to 1000 (the common first-user UID on single-user
+# Linux boxes) — override them with `id -u`/`id -g` if yours differ, so your host-native `pworkers`
+# CLI can still read them, e.g.: PW_UID=$(id -u) PW_GID=$(id -g) docker compose up -d research-desk
+#
+# Images are pinned by digest (tag kept for readability). Refresh a pin with
+#   docker buildx imagetools inspect <image>:<tag>   → the top-level "Digest:" line
 
 services:
   searxng:
-    image: searxng/searxng:latest
+    image: searxng/searxng:latest@sha256:85492ba800a861e3e8ff6014c430e81369fe6ca766a9990595f616d1ce95b272
     container_name: pw-searxng
     ports:
       - "127.0.0.1:8080:8080"
@@ -78,6 +81,9 @@ services:
       - PW_PORT=${PW_PORT:-8088}
       - PW_DB=/app/data/coordinator.db
     volumes:
+      # a fresh volume inherits the image's pw-owned /app/data. A volume created by an older,
+      # root-running image stays root-owned — hand it over once before upgrading:
+      #   docker compose run --rm --user root --entrypoint chown coordinator -R pw:pw /app/data
       - pw-coordinator-db:/app/data
     restart: unless-stopped