Disposable-host bundle-only end-to-end scenario campaign

[cbeaulieu-gt]

Part of #243 and the #361 release gate. Supersedes the GitHub-Actions execution proposed by #307.

Goal

Run one consolidated, real-daemon scenario campaign from the exact release bundle on a clean disposable Linux host that mirrors the supported deployment environment, without a Harness source checkout and without registering the host as a GitHub Actions runner.

Ownership boundary

External infrastructure:

• creates and destroys the clean disposable host;
• preinstalls required third-party prerequisites;
• creates and deletes the empty disposable sandbox repository;
• installs the dedicated GitHub App on that repository;
• supplies the operator, App-key-provider, executor, and Claude credentials.

Harness:

• verifies and installs the exact Publish a versioned wheel deployment bundle #382 bundle without dependency re-resolution or repository-source imports;
• provides working, non-authoritative example scripts for campaign setup;
• initializes scenario state through packaged application commands;
• runs and verifies all five scenarios;
• captures sanitized evidence and always attempts campaign-state cleanup.

The example scripts must not provision the host, install third-party prerequisites, create the repository, install the App, or contain application behavior unavailable through packaged bh-* commands.

Scenarios

Run the following against the externally supplied disposable sandbox:

1. terminal-block
2. recovery
3. clean-implement
4. block-ambiguity
5. ci-fail

No-agent and real-agent scenarios share the same campaign runner and evidence contract. Real-agent scenarios use the supplied Claude OAuth state; no mock-agent mode is permitted.

Evidence

Preserve, at minimum:

• release-bundle checksum;
• package version, exact source revision, and dependency-lock identity;
• host and prerequisite inventory;
• scenario reports and assertion results;
• command exit codes and sanitized logs;
• created issue, branch, and PR links;
• campaign cleanup outcome.

Setup, credential, GitHub-service, scenario-assertion, and cleanup failures must be distinguishable.

Acceptance criteria

• The campaign runs directly on the externally provisioned disposable host, with no GitHub Actions workflow or runner registration.
• The campaign uses only the supplied Publish a versioned wheel deployment bundle #382 release bundle; no Harness checkout or .git directory is present.
• A shipped example setup script works on a clean host whose prerequisites are already installed.
• The example setup script contains no unique application logic and does not cross the external-infrastructure ownership boundary.
• Packaged commands initialize, run, and verify all five scenarios against only the configured disposable repository.
• Operator, daemon-App, and executor identities remain separated throughout the campaign.
• All five scenario expectations pass against reports emitted by the real daemon.
• Evidence records the exact tested artifact and all fields listed above without leaking secrets.
• Cleanup is attempted after success or failure, and its outcome is recorded.
• External infrastructure can delete the host and sandbox after evidence capture without losing any required campaign artifact.

Dependencies

Depends on #359 and #382. #382 transitively supplies the packaged setup, initialization, verification, and application commands from #378–#381 plus the classified example assets from #383.

Out of scope

Host creation/destruction, operating-system provisioning, third-party prerequisite installation, sandbox-repository creation/deletion, GitHub App installation, credential enrollment, production activation, and rollback.

🤖 Generated by Codex on behalf of @cbeaulieu-gt