railiance-infra/history/2026-08-23-scope-against-intent.md

# Scope Against Intent Assessment — 2026-08-23

## Purpose

This assessment compares the repository's executable state on 2026-08-23 with
the stable direction in `INTENT.md`. It describes gaps; it does not create or
authorize implementation work.

The assessment was made after confirming that the repository had no actionable
local work: no ready, active, or blocked workplan; no task assigned to
`railiance-infra` in `todo`, `progress`, or `wait`; no unread inbox message; no
active repo dispatch; and no open Forgejo issue or pull request.

## Overall Assessment

The repository is a useful host-convergence and verification substrate, but it
does not yet meet the full intent of reproducibly turning bare machines into a
verified handoff. Its strongest implemented area is post-provisioning host
control. Its largest gap is the seam before that: neither current Host Europe
server is provisioned from source, and the only Terraform module cannot consume
the current inventory.

The intent is therefore **partially realized**. The repo can maintain important
parts of S1 for existing hosts; it cannot currently rebuild the live S1 estate
from its declared source alone.

## Principle-by-Principle Assessment

| Intent principle | Assessment | Current evidence | Gap |
| --- | --- | --- | --- |
| Declarative and reproducible | Partial | Ansible roles, firewall variables, a baseline spec, dynamic inventory, and a Hetzner Terraform module are source-controlled. | The live hosts are adopted Host Europe resources. The Terraform module iterates every inventory entry but requires `type`, `region`, `image`, and `role`, which the current entries lack. Host Europe lifecycle is manual, and no source path can recreate the current estate end to end. |
| Hardened by default | Partial | SSH password/root login are disabled by Ansible; fail2ban and UFW policy exist; k3s API access is tunnel-only; public web ports require a reef grant. | Hardening is applied after access exists rather than guaranteed as one provisioning transaction. `CoulombCore` deliberately bypasses UFW management. The generic cloud-init template still contains a placeholder SSH key and is not the Terraform template. |
| Verified before handoff | Partial | Goss renders inventory-aware checks, emits TAP, and runs hourly plus five minutes after boot. | The uniform baseline intentionally fails on `CoulombCore` because UFW is inactive. Verification is not a mandatory provisioning/handoff gate, and no repository CI exercises the Ansible/Goss contract. |
| Recorded source of truth | Partial | `inventory/servers.yaml` records the two names, IPs, and SSH users; host variables record the CoulombCore exception; Railiance01 has a non-secret evidence interface. | The inventory is authoritative for connection identity but not full resource shape or lifecycle. Current-host provider, location, flavor, image, and role are not modeled there. Evidence coverage is asymmetric and point-in-time. |
| Secure at rest | Partial | The Hetzner token is SOPS/age encrypted and repo hooks check files under `secrets/`. The host role installs SOPS/age without placing the private age key. | Rotation is manual. `inventory/group_vars/secrets.sops.yaml` is a plaintext placeholder outside the hook's checked path, while the docs describe it as encrypted input. The implementation and documentation disagree about the provider-token source. |
| Foundation, not tenant | Mostly met | Active convergence, firewall, SSH, inventory, and verification code stay at host level; cluster and platform ownership are documented elsewhere. | Historical Forgejo migration, runner, and restore-drill artifacts remain in the repo and blur discovery, although they do not form the current S1 execution path. |

## Direction-of-Evolution Assessment

### Stronger reproducibility and drift detection — partial

The hourly Goss timer is a real drift detector and its firewall assertions are
derived from the same inventory variables used for convergence. It records a
host flag and journald event, with an optional State Hub progress post.

Gaps:

- No central reaction is guaranteed; paging, task creation, and rollout
  blocking belong elsewhere and are not wired as a dependable end-to-end loop.
- `spec/server-baseline.yaml` is described as authoritative but is not consumed
  by Ansible or Goss. Maintaining the spec, roles, and template remains a manual
  three-way synchronization task.
- The declared spec and convergence are already imperfectly aligned: for
  example, the spec requires `htop` and a `tegwick` passwordless-sudo posture,
  while the base role does not establish both of those properties.
- A known expected failure means the all-host signal cannot distinguish only
  new regressions without host-profile semantics.

### Broader provider support — gap

Only Hetzner has Terraform resources. Host Europe is represented by adopted
identity/evidence and manual lifecycle notes, not a provider implementation.
The current single inventory cannot safely mix adopted Host Europe records with
the Hetzner `for_each` model.

### Continuous baseline verification — partial

The on-host hourly timer satisfies the cadence part of the direction. It is not
yet a complete continuous assurance capability because deployment does not
require a fresh green result, the failure route is optional, and one managed
host is expected to remain red.

### Automated rotation of at-rest secret material — gap

The repository provides `sops --rotate` as an operator command. It has no
scheduled or policy-driven rotation, recipient-age enforcement, rotation
receipt, or consumer rollout verification.

### Self-evidencing, auditable provisioning — partial

Goss TAP reports, capacity observations, ADRs, and workplan evidence are useful
records. The actual create/apply/converge sequence has no unified immutable
receipt tying together source revision, provider plan, created resource,
cloud-init completion, Ansible result, and a fresh verification result.

The local S1 backup command encrypts selected `/etc` files and package
selections, but it is not scheduled, does not copy off-host, and has no restore
procedure or drill in this repo. It should not be treated as proof that the
substrate is recoverable.

## Concrete Gaps, Ordered by Leverage

1. **Separate adopted and provisionable inventory semantics.** Add an explicit
   provider/lifecycle mode and validate the schema, or give each provider a
   filtered declaration. A Terraform plan must never interpret an adopted Host
   Europe entry as a Hetzner resource.
2. **Make one baseline contract executable.** Generate convergence and checks
   from a shared model, or add automated tests that fail when the human spec,
   Ansible roles, and Goss assertions diverge.
3. **Define host profiles and a real handoff gate.** Express the CoulombCore
   firewall exception in verification rather than accepting a permanently red
   host, and require a fresh result before S2 handoff.
4. **Repair the secret-source contract.** Remove or encrypt the plaintext
   placeholder file, extend plaintext checks to every declared secret path, and
   make the documentation name the provider-token source actually used by the
   Makefile.
5. **Capture provisioning receipts.** Record source revision, plan/apply
   identity, provider resource identifiers, convergence result, and verification
   result without recording secrets.
6. **Close recovery and rotation loops.** Automate age-recipient rotation with
   evidence, and give the local S1 backup a scheduled off-host copy plus a
   tested restore procedure if live-state backup remains part of this repo.
7. **Reduce scope noise.** Archive or relocate historical Forgejo/platform
   artifacts so current S1 ownership is evident from the file tree as well as
   from `SCOPE.md`.

## Bottom Line

Today the repo can answer: **"How do we keep these existing Linux hosts closer
to a hardened, observable S1 baseline?"**

It cannot yet answer the full guiding question in `INTENT.md`: **"Can we rebuild
the live substrate from source and prove, through a mandatory auditable gate,
that it is ready before higher layers use it?"**