Release v2.2.3 · buildcage/docker

Breaking Changes

Rename the sandbox action to run (#166)
- dash14/buildcage/sandbox@... doesn't describe what the action does — it applies buildcage's network-isolation technology to run an arbitrary command, not "sandbox" it in some more general sense. The action is now dash14/buildcage/run, matching what a workflow author actually calls it for. No backward-compatible alias is kept.
Before (v2.2.x):
```
- uses: dash14/buildcage/sandbox@<sha>
  with:
    run: npm install
```
After (v2.2.3):
```
- uses: dash14/buildcage/run@<sha>
  with:
    run: npm install
```

Generate a restrict-mode example in the run action's audit report (#170)
- The report action's audit-mode Job Summary already includes a ready-to-paste restrict-mode YAML example built from the audited hosts, saving a workflow author from hand-copying hostnames out of the table. The run action's own report renderer now generates the same kind of example — a run@<ref> step with proxy_mode: restrict and the discovered allowlist rules, plus the original run: command preserved so it stays copy-pasteable on its own

Handle read-only remount rejections on Docker-masked /proc mounts (#167)
- run-isolated.sh's read-only pass walks every mount under /proc/self/mountinfo and fails closed if a remount is rejected. Under Docker Desktop, /proc/scsi and /proc/interrupts are masked before the sandbox re-parents them under a fresh procfs, and the kernel then rejects an otherwise-harmless remount on them, breaking the mac dev loop without indicating any actual isolation gap. Mounts already reported read-only are now skipped, and remaining rejections get one retry re-bound onto themselves
Fix run's actionRef/actionRepo fallback for local-path invocations (#171)
- run/src/main.js fell through to an empty string instead of report/src/main.js's "v2"/"dash14/buildcage" defaults when invoked via uses: ./run, rendering a broken uses: /run@ in the generated audit-mode example
Fix a trailing blank line in the generated restrict-mode example's run: block (#172)
- GitHub Actions' run: | block scalar always keeps one trailing newline, which split() turned into a spurious empty line
Fix parseVertexAllowedLog to handle buildctl's multi-document JSON output (#172)
- A single JSON.parse() on the whole log text threw whenever buildctl debug logs --progress=rawjson flushed more than one newline-separated JSON document, silently dropping the explicit engine's Audited Hosts table and Communication details section

Add a label input to the run action (#172)
- Aligns its Job Summary heading (## Outbound Traffic Report — <label> (mode)) and Audited/Blocked/Allowed Hosts sub-headings with the report action's, so multiple run steps in the same job can be told apart in the summary

Add example-explicit.yml, an example workflow for the experimental explicit proxy engine (#172)
Add example-run-audit.yml, an example workflow for the run action's audit mode (#172)

Split test.yml into test-unit.yml, test-integration.yml, and test-e2e.yml (#157)
- Each keeps its own path-filter job and labeled job names (unit: test, integration: test (...), e2e: test (...)), replacing the single all-tests-passed required check with one aggregator per tier
Split the run action's e2e tests by case and consolidate its integration tests (#171)
- writable:/privilege-drop/filesystem-default cases and direct concurrent proxy lifecycles don't depend on the action wrapper, so they've moved out of test-e2e.yml into real shell scripts under run/test/, run via a single make test_sandbox_integration target; a new test_sandbox_audit_mode e2e job covers the restrict-mode example added in #170
Apply the latest Docker tag on release publish, not draft creation (#169)
- docker-publish.yml runs on tag push, which happens while creating the draft release, so latest was moving onto a new version before anyone reviewed or published it. Tagging latest is now handled by update-major-tag.yml, gated on release: published, alongside the existing major/minor floating tags

Move each action's Docker build context and test fixtures under its own directory (#160)
- setup/docker/, sandbox/docker/ (now run/docker/), setup/test/, sandbox/test/ were shared across actions despite belonging to only one of them, making ownership unclear as the actions diverged. Code actually shared across actions (image verification, Sigstore, rule/log parsing) was consolidated into a new core/

Full Changelog: v2.2.2...v2.2.3