Release-system reconciliation after the runner-host failure#

For the cross-document forensic analysis of why this design still failed to produce drastic simplification, start at ../forensics/2026-08-11-release-simplification-failure/README.md.

Status: source gate. No production mutation is authorized by this document.

Decision#

The surviving release system remains:

exact main event
  -> affected project build request
  -> immutable zero-traffic provider version
  -> ready evidence
  -> one environment/project publication lane
  -> expected-current read/check/publish/read
  -> generic health
  -> receipt or guarded restore

There is no global release, controller attempt, cumulative production baseline, runner target pointer, or host/application publication lane. Cloudflare's active stable Worker deployment is runtime truth. D1 is a build, evidence, attempt, and receipt read model.

What the last failure proved#

The private runner-host bootstrap created a Worker through the Beta Worker resource endpoint and then tried to upload its version through the legacy script-name endpoint. The resulting authorization failure was not evidence that the token merely needed broader permissions. It exposed a mixed provider model:

POST /workers/workers
POST /workers/scripts/{name}/versions

Cloudflare documents the resource-oriented Beta API as a coherent Worker and version model:

POST /workers/workers
POST /workers/workers/{worker_id}/versions
GET  /workers/workers/{worker_id}/versions/{version_id}

The private host will use that model consistently. The ordinary script-name version/deployment API remains the supported model for stable public Workers. The implementation must not bridge the two by name or widen a token until the old call succeeds.

Before another production attempt, a disposable, attempt-named Worker must prove the exact Beta create-version request, deploy semantics, declarative Durable Object/container export, provider-created namespace, private surface, and readback. The disposable proof must be deleted or retained as explicit canary evidence; it may not become runtime state or a resumable controller.

Project-scoped ordering#

Eligibility is scoped to (environment, project).

• A generation matters to a project only when that generation requested a build for that project.
• The highest eligible ready generation in that lane wins, regardless of build completion order.
• A newer failed build does not publish and does not change traffic.
• A late older ready build remains truthful evidence. It may publish only while no newer generation for that project has published or is the selected ready row; after a newer publication it becomes explicitly superseded.
• A generation affecting runner does not make an older deploy-control publication stale. Each project's receipt exposes its own source SHA and generation.

The repository-wide MAX(generation) used to allocate the next main generation is source ordering, not production eligibility. Publication queries must always include environment and project.

Expected-current boundary#

Build evidence may record the current stable deployment observed when the build finished, but that observation is not the publication precondition. The publisher captures and rechecks expected-current only after it owns the project lane and immediately before mutation. That same observed deployment and version are the rollback target. A dashboard or foreign API mutation causes a conflict and halts the lane; it is never overwritten.

Ready evidence, publication attempts, and terminal receipts are distinct:

• ready evidence proves one immutable zero-traffic version and its bindings;
• a publication attempt records fence, immediate expected-current, provider calls, readbacks, and health observations;
• a terminal receipt records published, failed-without-traffic, restored, or unresolved outcome without rewriting prior evidence.

Tables introduced without D1 triggers are application-enforced append-only. No documentation may claim that D1 itself rejects every update or delete unless the deployed schema demonstrably does so.

Normal main lifecycle acceptance#

The cutover is not normal until one ordinary merged PR performs all of the following without seed-generation, manually exchanged SHAs, provider IDs, fences, or lane-clear messages:

1. required Platform CI and complete Platform Preview checks pass;
2. main CAS appends one immutable generation;
3. the repository event requests builds only for affected projects;
4. no legacy release attempt, production candidate, operation/component, or controller wake/adoption row is created;
5. the builder obtains the registry marker and trusted egress replaces it with a short-lived upstream credential without persisting or logging the secret;
6. the builder writes ready evidence;
7. the per-project publisher selects, publishes, verifies, and receipts it.

The first real proof remains Better Auth plus NOMOREGIT invite flow. Two more independent feature releases follow through the same path.

Token roles#

Partial artifacts and cleanup#

Every attempt-derived artifact set is classified as building, ready, failed, superseded, or published. A failed set records every provider object known to exist and the operation that failed. Existing matching write-once objects may be adopted only after exact readback; partial, ambiguous, or mismatching sets fail closed and are never repaired in place.

Garbage collection is separate from build, publish, restore, and bootstrap. It may delete only terminal, never-published objects after a retention interval and after proving they are not referenced by current traffic, rollback targets, ready rows, in-flight attempts, or unresolved evidence. The historical local runner container remains untouched through bootstrap and three normal proofs.

Legacy and bootstrap deletion gates#

Before calling the system simplified, prove by code search, tests, runtime routes, bindings, and database deltas that the normal event cannot enqueue, wake, adopt, resume, reconcile, or retry the legacy controller system. Historical product migrations may remain for protected production data, but legacy release authority must be unreachable.

After the three normal proofs and rollback-compatibility window:

• delete or permanently disable bootstrap-production-release-cutover.mjs;
• delete manual generation seeding;
• delete the synthetic preview-cutover state machine after the two required checks are ordinary configuration;
• delete remaining legacy release/controller runtime paths and bindings;
• schedule removal of the old local runner container as a separate cleanup.

Stop conditions for the next attempt#

No production mutation occurs until all are true:

• disposable Beta Worker/version/container proof passes from provider readback;
• the runner builder uses one resource family consistently;
• reverse-completion and per-project independence tests pass;
• an ordinary main event creates project build requests with zero legacy rows;
• token roles are independently configured and read back;
• the exact prior stable deployment/version and data rollback evidence are recorded;
• one ten-minute, one-attempt bootstrap checklist is prepared from clean exact main.