fix: preserve Settings SSH policy and prevent gateway resize crashes

[meffmadd]

Summary

Preserve the gateway guarantees needed to test the Settings POC with this fork image, and fix the resize/exit crash observed during manual testing.

 SSH session channel
+  accept only one successful shell/exec/subsystem start (failed starts remain retryable)
 signal request
-  consult Shell.Mode
+  consult Signal.Mode
 connection-pod readiness failure
+  roll back the newly created pod with an independent, bounded cleanup context
 agent PID before initial PTY resize
-  wait for six bytes for every PTY
+  read four bytes for connection/persistent exec; retain session PTY wait/echo framing
 execution finishes → resize queue Stop
-  close the data channel while window requests may still send
+  signal shutdown separately and idempotently; release blocked readers/writers

A late window-change previously caused panic: send on closed channel, crashed the entire gateway, broke the port-forward, and stranded connection-mode Settings pods. The resize regression exercises real execution completion and OnWindow in all three Kubernetes modes.

The other four fixes were already required by the deployed dev.box gateway. They are now ported into this PR with their existing regressions so replacing that gateway with the fork image does not reintroduce startup deadlock or weaken Settings policy/lifecycle behavior. The user explicitly approved this expanded scope.

CI publication now compares the PR head repository with github.repository, rather than testing whether the repository itself is a fork. Same-repository PRs can publish pr-11; external PRs remain excluded. Push, attestation, and summary conditions match. The earlier successful image build used push: false and did not publish its tag.

Target: TU-Wien-dataLAB/ContainerSSH → deploy, not upstream. No chart/webhook/application code, orphan GC, or automatic cluster deployment is included.

Evidence

• Before: fork baseline reproduces concurrent/post-exit OnWindow panics and repeated-Stop panic. Each of the four added regressions also failed before its production fix: repeated starts accepted, signal policy routed through Shell.Mode, no readiness rollback, and four-byte PID blocked before initial resize.
  After: all five regressions pass with race detection:

  go test -race -mod=readonly ./internal/sshserver ./internal/security ./internal/kubernetes \
    -run '^(TestSettings|TestResizeQueueShutdown)' -count=20 -timeout=90s

  The resize regression also passed 50 race-enabled repetitions independently.
• Native gateway build, targeted vet, and golangci-lint 2.14.0 (diff against origin/deploy) pass with zero issues.
• Actual fork merger, security/pod validators, and native binary parse/defaults accept the existing release base plus effective Settings and ordinary-box configurations.
• Publication conditions cover pushes, manual runs, same-repository PRs, and rejection of external PR publication. Standards/spec review found no blockers.
• The unrelated S3 image failure is now fixed in the deploy base by upstream PR #713, which switches to pgsty/minio. Merged updated deploy into this branch and reverted the temporary source-built fixture. S3 helpers, upload tests, and their READMEs now match deploy exactly; they are not part of this PR's diff.
• Full CI green at 5f6c644: pull-request tests and push tests pass. Lint, both snapshot builds, and image publication also pass; the push snapshot succeeded on rerun after an external release-query HTTP 403. Targeted vet/lint and all Settings/resize regressions pass locally with race detection for 20 repetitions. Local Docker integration reruns remain blocked by a full Docker VM disk.
• New PR image published: ghcr.io/tu-wien-datalab/containerssh:pr-11@sha256:a2e40e8265787a86b5b49579dd4fc7fa7515641b244121d7e5d3e2104047914c. This image has not been rolled out or live-tested; the live evidence below refers to the earlier pinned digest.
• User-approved live test passed (2026-10-14): dev.box release revision 21 uses ghcr.io/tu-wien-datalab/containerssh:pr-11@sha256:529a4f5618be89a56e3714fe9f77cf886a6674131e0cec5f5f1b28841d92beb9. Gateway-only upgrade preserved all other values/Secret references. Full SSH policy/lifecycle probe, 112 quit/resize packets across four fresh connections, and real OpenSSH checks passed; gateway UID remained stable with zero restarts, new Settings pods were removed after disconnect, and the ordinary box UID was unchanged. Two pre-existing crash leftovers were deliberately not deleted. Live evidence.

Merge Danger

Door: two-way

No configuration/API migration or persistent-data change. Reverting the commits restores prior behavior, including the bugs. A resize already in flight may still reach an active reader while shutdown races it; calls starting after shutdown are rejected. Readiness rollback bounds the DELETE request, not finalizer resolution or physical disappearance.

Blast Radius: Gateway

Single-start enforcement affects SSH session channels across backends; signal policy follows its own configuration. Kubernetes PID/resize behavior changes across execution modes, while new readiness rollback is connection-mode only. CI can publish trusted same-repository PR images. Live image selection and rollout are separate operations; existing orphaned Settings pods still require separately approved, UID-preconditioned cleanup.