Summary

TEM-27. Removes the unused FileBrowser binary from runpod/base, patches the
outstanding CRITICAL/HIGH vulnerabilities across our image families, turns Grype
into a hard gate on the families we've cleaned up, and hardens the build/release
pipeline so flaky BuildKit cache failures can no longer silently skip a release
or mis-bump the version.

FileBrowser is being archived on Sept 1st
by upstream and ships with known, will-not-fix security advisories. We installed
the binary in base but never started it (only comfyui-base runs it), so it
was pure attack surface with no benefit — dropped it.

Changes

FileBrowser

• base: remove the FileBrowser install step entirely (binary was never
  launched by start.sh, so no runtime behavior changes).

Vulnerability fixes

• pytorch: re-install a patched Pillow from PyPI over the CVE-affected copy
  torchvision drags in from the PyTorch wheel index. Version is centralized as
  PILLOW_VERSION in docker-bake.hcl and only applied when Pillow is actually
  present (torch-only builds are untouched).
• pytorch-cluster: bump node_exporter 1.11.1 → 1.12.1 and Grafana
  13.1.0 → 13.1.1 (Go stdlib / dependency CVEs in the embedded binaries).
• autoresearch: bump runpodctl v2.3.0 → v2.7.2 (Go stdlib / x-crypto CVEs).
• nvidia-pytorch: jupyterlab 4.5.9 → 4.5.10, mistune 3.2.1 → 3.3.0,
  add msgpack==1.2.1, add soupsieve==2.8.4, pillow 12.2.0 → 12.3.0.
• rocm: jupyterlab 4.5.9 → 4.5.10, add msgpack==1.2.1,
  pillow 12.2.0 → 12.3.0.
• base: jupyterlab 4.5.9 → 4.5.10.

CI - vulnerability gating

• grype: add a fail-on-findings input (default false, report-only).
  Enabled (true) for the cleaned-up families: base, pytorch, nvidia.
  Other families stay report-only until their remaining findings are resolved.

CI - build robustness

• docker-setup: upgrade useblacksmith/setup-docker-builder v1 → v2.0.1 and
  add a required cache-key input. Each image family now uses a distinct cache
  lineage (<repo>/base, /pytorch, /cluster, /nvidia, /rocm,
  /autoresearch) so families no longer thrash each other's BuildKit sticky-disk
  cache — this also fixes the "no space left on device" failures on ROCm builds.
• bake-retry (new composite action): every docker buildx bake step now runs
  through .github/actions/bake-retry. On a transient BuildKit failure (typically
  sticky-disk cache corruption: failed to compute cache key … unexpected commit digest … failed precondition) it prunes the builder cache and retries once with
  --no-cache, so a poisoned cache entry self-heals within the run instead of
  turning the build red. Centralised so each bake step stays a one-line call.

CI - release correctness

• compute-version: on push/release, detect the Conventional-Commits bump from
  the merge commit's subject only, not its body or the whole
  last-tag..HEAD range. A ci:/chore: squash merge whose body lists
  feat:/fix: branch commits (or that follows earlier un-released commits in
  the range) no longer triggers a phantom version bump.
• manual-release (new break-glass workflow): workflow_dispatch that creates
  the git tag + GitHub Release for an already-built-and-pushed version, without
  rebuilding. Recovers the case where a flaky build left release stuck in
  skipped (GitHub's "Re-run failed jobs" never resurrects a skipped downstream
  job). Guards: main-only, tag-format validation, refuses if the tag already
  exists, and shares the push-release concurrency lane.

Notes / follow-ups

• pytorch-cluster (dcgm-exporter), rocm (base OpenSSL/Python), and
  autoresearch (pillow/urllib3 via its own uv.lock) still have findings
  that depend on upstream releases — those families remain report-only for now.