test(e2e): full seal suite vs the dev backend (backend#1184 fast-follow)

[LukasWodka]

What (the backend#1184 deferred fast-follow)

backend#1184 closed accepting CI-substrate coverage as v1-sufficient, with one recorded residual: "the FULL suite (incl. backend-reachability + bound-PVC storage-assertions) on the dev-env e2e-test-agent with real credentials." This is that harness.

scripts/tests/e2e-full-seal.sh (new)

Real k3d cluster → install the working-tree chart as the dedicated dev e2e-test-agent client (real creds, CLIENT_ENV=dev, lockdown engaged, same base profile as the sibling) → crisp pre-asserts (every release PVC Bound within 300s — separates "slow to bind" from "assertions wrong"; mysql + jobs-manager rolled out, i.e. a real backend session) → shared egress positive control → hook-presence guard for all three checks (an unfiltered helm test "passes" a release whose hooks silently stopped rendering — same trap the sibling guards for its --filter) → helm test unfiltered: egress-enforcement + backend-reachability + storage-assertions in one release. On failure, every seal-check pod log dumps via the §8.2 enumeration label.

full-seal-e2e job (helm-ci)

• push / workflow_dispatch only, never PRs — fork PRs can't read secrets, and a real dev-backend login per PR is platform churn for no extra signal (the secret-free enforcement probe keeps covering PRs).
• Skips green with a ::notice until provisioned — the job is live the moment the secrets exist, no second PR needed.
• Job-level concurrency (one dev-agent session at a time — two clusters on one client id reads as a capacity anomaly).
• workflow_dispatch added to helm-ci for on-demand runs (e.g. right after provisioning).

Refactor + docs

• The egress positive control moved verbatim from e2e-seal-check.sh into lib/e2e-common.sh (provenance + test(seal-check): run the egress-enforcement probe live in k3d CI (#1184) #541 review rationale documented at the function) — one copy, two callers, no drift.
• SEAL-CHECK.md: new CI-coverage map (what runs where, activation contract, "never a person's shared dev identity — the backend#1180 token-churn failure class"), and the follow-ups list drops three items that shipped since it was written (test(seal-check): run the egress-enforcement probe live in k3d CI (#1184) #541 live probe, cli#393 CLI verdict, cli#449 §8.3 matrix).
• Both installer-tests.yaml shellcheck gates now include the new script; helm-ci path filters gain the new script + lib/e2e-common.sh (which was already load-bearing for the seal job but absent from the filter).

⚠️ Activation — the one human step (not in this PR)

Provision a dedicated dev-platform client (e2e-test-agent) and add two repo Actions secrets: TB_E2E_CLIENT_ID, TB_E2E_CLIENT_PASSWORD. Until then the job skips green by design. (Public repo: fork PRs can't read the secrets; the job never runs on PRs anyway.)

Test plan

• bash -n + shellcheck --severity=error green on all three touched scripts (both CI shellcheck lists updated).
• Workflow YAML parse-validated; R8 manifest untouched by design (scripts/tests/ isn't manifest-covered); style guard clean.
• The refactored sibling (seal-check-e2e) runs in THIS PR's CI — proving the factored positive control live on k3d.
• The new job's full path can't run pre-provisioning by construction; first green run gets recorded in SEAL-CHECK.md per the doc.

Epic: tracebloc/backend#1151 (workstream G residual) · closes the backend#1184 fast-follow · RFC-CLI-0003 D12 §8.2

🤖 Generated with Claude Code

---

Note

Low Risk
Changes are CI/e2e harness and documentation only; no chart runtime or production install path. Real credentials stay in Actions secrets and a dedicated dev test client, with PRs excluded from the credentialed job.

Overview
Adds e2e-full-seal.sh and a helm-ci full-seal-e2e job to run the complete chart seal suite (egress-enforcement, backend-reachability, storage-assertions) on k3d against the dev backend with dedicated e2e-test-agent credentials—not just the secret-free egress probe PRs already get.

The new script installs the working-tree chart with real creds in a mode-0600 values file, waits for PVCs to bind and for jobs-manager rollout, verifies all three test hooks exist (guards against silent empty helm test), then runs unfiltered helm test with failure log dumps via the seal-check label.

CI wiring: workflow_dispatch on helm-ci; path filters include the new script and e2e-common.sh; installer shellcheck lists updated. The job runs only on push/dispatch (not PRs), uses 45m timeout and concurrency so one dev-agent session runs at a time, and skips green with a notice until TB_E2E_CLIENT_ID / TB_E2E_CLIENT_PASSWORD secrets exist.

Refactor: egress positive-control logic moves from e2e-seal-check.sh into e2e_egress_positive_control in e2e-common.sh for both seal e2e scripts.

Docs: SEAL-CHECK.md gains a CI coverage map and trims follow-ups that shipped elsewhere.

^{Reviewed by Cursor Bugbot for commit d639aca. Bugbot is set up for automated code reviews on this repo. Configure here.}

[LukasWodka]

bugbot run

[LukasWodka]

All three findings addressed: creds now travel in a 0600 temp values file (never argv — and immune to helm --set's comma/brace mangling), the PVC wait retries through kubectl blips instead of aborting under set -euo pipefail, and the job cap is 45m with the summed script bounds documented.

bugbot run

[LukasWodka]

Second-round finding fixed: the install now pins pvcAccessMode=ReadWriteOnce (chart default is RWX, which local-path never provisions — the Bound pre-wait and storage-assertions were structurally unable to pass; installer parity for this storage path).

bugbot run

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit d639aca. Configure here.}