fix: eliminate Docker test flakiness and cut CI time

[marconae]

Summary

• Cache Docker image in GitHub Actions (type=gha) so apt-get + Claude CLI + Codex CLI installs are not re-downloaded on every run — image build becomes a near-instant cache hit after the first push
• Parallelize all 4 integration test containers — each is fully isolated, so wall-clock time drops from the sum of all four to the slowest single container
• Skip redundant docker build in test-docker.sh when the image is already loaded by the CI pre-build step; local dev still builds on demand
• Remove || true on Claude CLI install in Dockerfile so a failed install surfaces immediately instead of silently producing a broken test environment
• Drop the 3-attempt retry loop — the primary flakiness source was the fresh Docker image pull on every run, which caching eliminates

Root causes addressed

Source	Type	Impact
Fresh Docker image pull (apt + Claude CLI + Codex CLI) on every run	Slowness	8–12 min saved per run
4 sequential container runs	Slowness	3–5 min saved by parallelizing
Retry re-ran full pipeline including slow Docker build	Slowness + waste	Eliminated
|| true on Claude CLI install	Silent flakiness	Fails explicitly now

Expected CI times

Scenario	Before	After
Cache hit (no Dockerfile change)	30 min	~10 min
Dockerfile changed (cold cache)	30 min	~18 min

Test plan

• CI run on this branch completes in < 15 min
• A second push (no Dockerfile change) shows a GHA cache hit on the Docker build step and completes in < 12 min
• All 4 test suites (install, codex-plugin, update, uninstall) still pass