e2e-verify `tracebloc delete` offboard teardown against a throwaway k3d cluster

[saadqbal]

Follow-up to #139 / #136. Offboarding — the top-level tracebloc delete (deprovision/revoke + helm uninstall + k3d cluster delete + clear the active-client pointer) — is currently unit-tested only (fake runner in internal/nodeboot; logic in internal/cli/delete.go). CI e2e.yml covers data ingest, not teardown. Before the offboarding surface promotes to prod, run it end-to-end on a throwaway k3d box:

• create a client, confirm the release + cluster exist;
• tracebloc delete --yes;
• confirm: the backend credential is revoked (POST /edge-device/<id>/revoke/ — not a hard DELETE; the client row is kept as history per §7.10), the Helm release is uninstalled, the k3d cluster is deleted, ~/.tracebloc is wiped (or kept under --keep-data with the pointer cleared), and a foreign tb alias / system tooling is left untouched (the fix(cli): delete only removes tb when it is tracebloc's own alias #171 guard).

Do NOT run against the shared reference box (18.196.47.80) — it would destroy the live tracebloc-amazon client.

Rescoped 2026-07-07: client delete → the shipped top-level tracebloc delete; the teardown surface landed in #160/#165/#168/#169/#171 (now Ready for staging), so this is the pre-prod e2e gate.

[aptracebloc]

e2e teardown gate implemented — closing.

Delivered by #236 (commit 78ed06d, merged 2026-07-13). A new delete-teardown job in .github/workflows/e2e.yml builds the real tracebloc binary and runs the offboard black-box against a throwaway k3d cluster (k3d, not kind, because teardown shells out to k3d cluster delete). Tests in test/integration/delete_e2e_test.go:

• TestE2E_RevokeUsesPostNotDelete — asserts the credential offboard hits POST /edge-device/<id>/revoke/ and never a hard DELETE (row kept as history per §7.10).
• TestE2E_DeleteTeardown — tracebloc delete --yes --force against a real Helm release on a throwaway k3d cluster; asserts the release is uninstalled, the cluster is deleted, ~/.tracebloc is wiped, and a foreign tb alias is left untouched (fix(cli): delete only removes tb when it is tracebloc's own alias #171 guard). Opt-in via TB_E2E_K3D=1 (CI sets it) so it can't clobber a dev machine's cluster.

One documented caveat (called out in test/integration/README.md, not a regression): CI exercises the revoke against a stub backend (best-effort transport path); the POST-not-DELETE contract is pinned by the stub test, but the revoke against a live backend (server-side row-preserving behavior) is intentionally deferred to the pre-prod FR gate — matching this issue's own warning not to run against the shared reference box.