informed-decision/deploy/lifecycle/RELEASE-INFD-WP-0010-T02.md

# Release runbook: present_existing for nk-bind-executor + compact-sitting v14 (INFD-WP-0010-T02)

Prepared 2026-11-01 by the informed-decision owner agent. **Nothing in this file has been applied.**
Founder approval (2026-11-01, relayed by the net-kingdom coordinator): release present_existing together with
access-engine compact-sitting v14. Every command below is founder-run, with
`export KUBECONFIG=$HOME/.kube/config-railiance01`.

## Status: blocked on the service image (gate G0)

No CI image exists for current `main`. This repository has no Forgejo workflow that builds the full service
image. Its two workflows build single-purpose overlays on fixed base digests:

- `.forgejo/workflows/purpose-overlay-image.yaml` swaps one module, `policy.py`, on base `7f5c6cb2…`, and
  checks it against the pinned hash `9d7c1be4…`. Commit `80ef840` changed `policy.py` (now `912ad3a3…`), so
  run 5 (head `e617bb5`, 2026-11-01T03:07Z) failed at that check. That is the intended behaviour. Even if the
  hash were updated, the overlay could not carry this release: it touches nine modules (`intake.py`,
  `intake_service.py`, `intake_approval.py`, `outcome.py` (new), `policy.py`, `requester.py`, `review.py`,
  `store.py`, `workload.py`).
- `.forgejo/workflows/zlib-overlay-image.yaml` upgrades zlib only, on base `763c3640…`.

The registry has no tag for `3b7031a`, `main` or `latest`. `docker buildx imagetools inspect` returns
not found for each.

Before this release can go out, an owner-CI full-image build from the release source must exist. It needs
its own workflow, and it must keep zlib at `1.3.2-r1` or later. A plain rebuild of `Containerfile` uses the
pinned `python:3.12-alpine@sha256:b64631e0…` base and only adds `libuuid`. That would reintroduce the zlib
finding that the live image `f5588f49…` remediated. Then pin the image as in step 0.

**Update, same day:** the full-image workflow now exists in source: `.forgejo/workflows/service-image.yaml`. It
builds `Containerfile` from the pushed commit's tarball, with the wheel built in-image. It installs the overlay's
hash-pinned zlib-1.3.2-r1 package offline and enforces `zlib>=1.3.2-r1`. It checks the runtime smoke and that the
installed source matches the commit before pushing `forgejo.coulomb.social/coulomb/informed-decision:main-<short sha>`,
then prints the digest in "Report immutable digest". A local build of the same Containerfile passed 13/13 container
checks (`tools/smoke_container.py --workload`), with zlib 1.3.2-r1 and the installed modules identical to source.
The gate is now: the push runs this workflow, then step 0 pins the digest it reports.

**Do not sync this candidate with the current image.** The live image `f5588f49…` accepts only the four basic
registry keys: `set(row) != {submitter_id, memo_classes, review_groups, daily_limit}`. It would reject the
executor row, so every intake request would fail with 503 `intake_configuration_unavailable`, State Hub's
included. It also does not send `approval_action`.

## Pins

| | Current, live (rollback target) | Candidate |
| --- | --- | --- |
| Argo `informed-decision` targetRevision | `312ab97e876e17341bf16f9e95397e14c2ac2720` | `<RELEASE_SHA>` (step 0) |
| Service image | `…/informed-decision@sha256:f5588f496b1f267d07cdc9f5208e6f2f241f74b739c23924b03702dec13a5c23` | `<IMAGE_DIGEST>` (G0) |
| Runtime ConfigMap | `informed-decision-runtime-6c221878d051` | `informed-decision-runtime-b488b96a0953` |
| Review policy pin | compact-sitting `v13` `sha256:31be5985c0c1e6b0c3a039ed7310d1e94daf523a94f6d926e6e8d759d3711873` | `v14` `sha256:8911458f36f1a60598aa19592a2d752a74dffe2617fd86238407ac87bc300fde` |
| Helm `flex-auth-informed-decision-sitting` (access-engine) | revision 13, image `6c6dc263…`, v13 policy path | access-engine's v14 release (`2de0f86`), prepared in parallel |
| Lifecycle policy pin | `informed-decision.lifecycle` v2 `da01e045…` | unchanged |
| Store schema | 5 | 5 (no migration) |

Candidate source changes in this commit:

- `admission.candidate.json`: the policy pin moves to v14, and the workload peer `{namespace: sso, app: bounded-executor}`
  is added. `service.candidate.json` is re-rendered with `tools/render_deployment.py`.
- `informed-decision-review` NetworkPolicy: the 8081 ingress rule `from` now lists exactly two peers:
  `state-hub/app.kubernetes.io/name=state-hub` (unchanged) and `kubernetes.io/metadata.name=sso` AND
  `app.kubernetes.io/name=bounded-executor`. No other sso pod is admitted, and nothing new is admitted on 8080.
- `intake-submitters.json` (projected into `registries.candidate.json`): new row
  `system:serviceaccount:sso:nk-bind-executor` → `nk-bind-executor`, memo class `bounded-execution`, review group
  `net-kingdom-admins`, `daily_limit: 10`, `present_existing {approval_actor: service:nk-bind-executor, actions:
  [bounded-execution]}`. The state-hub row is byte-identical, and `outcome_read` stays off (T05 waits for STATE-WP-0101).

## Step 0: pin the image (after G0; owner agent or founder)

```bash
cd ~/informed-decision
docker buildx imagetools inspect forgejo.coulomb.social/coulomb/informed-decision:main-<short sha>   # or the CI "Report immutable digest" line
# set "image" in deploy/lifecycle/admission.candidate.json to forgejo.coulomb.social/coulomb/informed-decision@<IMAGE_DIGEST>
python3 tools/render_deployment.py --input deploy/lifecycle/admission.candidate.json \
  --output deploy/lifecycle/service.candidate.json
.venv/bin/python -m pytest -q                       # expect 0 failures
python3 tools/smoke_container.py --image forgejo.coulomb.social/coulomb/informed-decision@<IMAGE_DIGEST> \
  --receipt docs/evidence/<date>-present-existing-release/container.json --workload
git commit -am "INFD-WP-0010-T02: pin service image <IMAGE_DIGEST> for the v14 release" && git push
RELEASE_SHA=$(git rev-parse HEAD)
```

Only the image line and the re-rendered `service.candidate.json` change. The runtime ConfigMap name stays
`…b488b96a0953` because the image is not part of runtime.json.

## Step 1: preflight (read-only)

```bash
export KUBECONFIG=$HOME/.kube/config-railiance01
kubectl -n argocd get application informed-decision \
  -o jsonpath='{.spec.source.targetRevision} {.status.sync.status} {.status.health.status}{"\n"}'
#   expect 312ab97e876e17341bf16f9e95397e14c2ac2720, Healthy (OutOfSync is the known orphan ConfigMap
#   informed-decision-runtime-c81577741a09; prune stays off)
kubectl -n informed-decision get deploy informed-decision -o jsonpath='{.spec.template.spec.containers[0].image}{"\n"}'
helm -n flex-auth history flex-auth-informed-decision-sitting --max 3     # expect revision 13 deployed
kubectl -n sso get sa nk-bind-executor
kubectl -n informed-decision exec deploy/informed-decision -- \
  informed-decision-admin inspect --db /data/private/review.sqlite        # record counts (schema 5)
# Server dry-run of the exact release source (no persistence):
git -C ~/informed-decision checkout --detach "$RELEASE_SHA"
kubectl apply --dry-run=server -k ~/informed-decision/deploy/lifecycle
kubectl diff -k ~/informed-decision/deploy/lifecycle    # expect: image, runtime ConfigMap name/new ConfigMap,
                                                       # registry ConfigMap, review NetworkPolicy (+ Argo annotations)
git -C ~/informed-decision checkout main
```

## Step 2: backup (schema is unchanged, but the release convention still requires it)

```bash
STAMP=$(date -u +%Y%m%dT%H%M%SZ)
kubectl -n informed-decision exec deploy/informed-decision -- \
  informed-decision-admin backup --db /data/private/review.sqlite --output /data/private/review-$STAMP-pre-v14.sqlite
mkdir -p -m 0700 ~/informed-decision/.private-backups/$STAMP-pre-v14
kubectl -n informed-decision exec deploy/informed-decision -- cat /data/private/review-$STAMP-pre-v14.sqlite \
  > ~/informed-decision/.private-backups/$STAMP-pre-v14/review.sqlite
chmod 0600 ~/informed-decision/.private-backups/$STAMP-pre-v14/review.sqlite
sha256sum ~/informed-decision/.private-backups/$STAMP-pre-v14/review.sqlite
```

## Step 3: one window, access-engine first, then informed-decision

The service checks the PDP's provenance on every review call: `policy_package`, `policy_version`,
`policy_package_digest` and `matched_policy_version` must all equal its pin. Between the two steps, the
version mismatch makes human review reads and acts **fail closed**. No review is lost, but pages refuse.
Workload intake (`POST /v1/memos`) does not consult the PDP, so State Hub posing continues throughout. Keep
the gap to minutes.

**3a. access-engine v14** follows access-engine's own release runbook (FLEX-WP-0040-T07,
`access-engine/docs/compact-sitting-v14-present-existing.md`, "Live steps"). It runs a Helm upgrade of
`flex-auth-informed-decision-sitting` with `--atomic`: a new image built from `2de0f86` or later, and
`--policy /opt/flex-auth/examples/informed-decision-present-existing-review/policy.md`. Check:

```bash
helm -n flex-auth history flex-auth-informed-decision-sitting --max 2      # expect revision 14 deployed
kubectl -n flex-auth rollout status deploy/flex-auth-informed-decision-sitting --timeout=180s
```

If the `--atomic` upgrade rolls itself back, stop here. Do not run 3b. Nothing on the informed-decision side has changed.

**3b. informed-decision**, through the manually pinned Argo application:

```bash
kubectl -n argocd patch application informed-decision --type merge \
  -p "{\"spec\":{\"source\":{\"targetRevision\":\"$RELEASE_SHA\"}}}"
kubectl -n argocd patch application informed-decision --type merge \
  -p "{\"operation\":{\"initiatedBy\":{\"username\":\"founder\"},\"sync\":{\"revision\":\"$RELEASE_SHA\",\"prune\":false,\"syncOptions\":[\"ApplyOutOfSyncOnly=true\"]}}}"
kubectl -n argocd get application informed-decision \
  -o jsonpath='{.status.operationState.phase} {.status.sync.revision} {.status.health.status}{"\n"}'
kubectl -n informed-decision rollout status deploy/informed-decision --timeout=300s   # Recreate: short restart
```

## Step 4: verification

Possible without the executor:

1. **Pins live.** Check the Deployment image and that its `configuration` volume names
   `informed-decision-runtime-b488b96a0953`:
   ```bash
   kubectl -n informed-decision get deploy informed-decision \
     -o jsonpath='{.spec.template.spec.containers[0].image} {.spec.template.spec.volumes[?(@.name=="configuration")].configMap.name}{"\n"}'
   kubectl -n informed-decision get networkpolicy informed-decision-review -o jsonpath='{.spec.ingress[1].from}{"\n"}'
   #   expect exactly the state-hub peer and the sso + app.kubernetes.io/name=bounded-executor peer
   ```
2. **Public surface.** `curl -s -o /dev/null -w '%{http_code}\n' https://decisions.coulomb.social/healthz` and
   `/readyz` return 200. `/v1/memos` and `/v1/decisions` on the public origin return 404.
3. **v14 pin match (human, aal2).** Open `https://decisions.coulomb.social/` and one existing memo you were
   presented. The page loads only if the PDP's v14 provenance equals the new pin. A refusal mentioning policy
   or provenance means 3a and 3b disagree: roll back (step 5).
4. **Store unchanged.** Run the `informed-decision-admin inspect` command from step 1 again. Schema 5, and
   the counts are no lower than the preflight.
5. **Executor path, synthetic and side-effect free.** Run this from a temporary pod that carries the
   executor's identity and label. It is a founder-run kubectl write: a short-lived pod in sso, removed by
   `--rm`. The probe pins a deliberately wrong binding digest, so even a successful approval read ends in a
   409. No memo is stored, no budget is spent, and nothing is created, entered or consumed in approval-engine.
   ```bash
   cd ~/informed-decision
   IMAGE=forgejo.coulomb.social/coulomb/informed-decision@<IMAGE_DIGEST>
   NAME=infd-probe-$(date -u +%H%M%S)   # --overrides (merge patch) replaces the generated container: it carries name and image
   OVR='{"spec":{"serviceAccountName":"nk-bind-executor","automountServiceAccountToken":false,
     "securityContext":{"runAsNonRoot":true,"runAsUser":10001,"seccompProfile":{"type":"RuntimeDefault"}},
     "containers":[{"name":"'$NAME'","image":"'$IMAGE'","command":["python","-"],"stdin":true,"stdinOnce":true,
       "securityContext":{"allowPrivilegeEscalation":false,"readOnlyRootFilesystem":true,"capabilities":{"drop":["ALL"]}},
       "volumeMounts":[{"name":"intake","mountPath":"/var/run/secrets/informed-decision-intake","readOnly":true}]}],
     "volumes":[{"name":"intake","projected":{"sources":[{"serviceAccountToken":{"path":"token",
       "audience":"informed-decision-intake","expirationSeconds":600}}]}}]}}'
   .venv/bin/python tools/present_existing_probe.py --approval-id "$(python3 -c 'import uuid;print(uuid.uuid4())')" --program \
   | kubectl -n sso run "$NAME" --rm -i --restart=Never --image="$IMAGE" \
       --labels=app.kubernetes.io/name=bounded-executor,app.kubernetes.io/component=release-probe --overrides="$OVR"
   ```
   Expected result for a random UUID: `503 {"error": "approval_unreadable"}`. Getting there proves the
   NetworkPolicy admits the sso executor label on 8081, TokenReview accepts `nk-bind-executor` with audience
   `informed-decision-intake`, the reloaded registry parses (so the state-hub row is also loaded), the
   `present_existing` profile and envelope validation pass, and the reservation is released.
   - If you know an approval id in approval-engine that no informed-decision memo has presented yet (for example
     one created during NK-WP-0056 executor testing), use it in place of the random UUID. Expect a 409 with an
     `approval_*` reason (`approval_digest_mismatch` or an earlier one). That proves the read through KeyCape
     `approval:read` as well.
   - Negative check: rerun with `"serviceAccountName":"default"` and expect `403 submitter_unknown`.
   - Other outcomes: `401 workload_authentication_failed` points to the token or audience. A timeout points to the
     NetworkPolicy (check the executor egress policy too). `503 intake_configuration_unavailable` means the
     registry did not parse: roll back.

Needs the executor (this is T02's own done-when): the executor creates an approval as `service:nk-bind-executor` and
presents it, the founder accepts at aal2 on `/review`, and the executor consumes it. That is the
NK-WP-0056 pilot run. It is not part of this release window.

## Step 5: rollback to the current release

The database is retained and stays on schema 5, so this is not a downgrade. Do both halves in the same window:

```bash
kubectl -n argocd patch application informed-decision --type merge \
  -p '{"spec":{"source":{"targetRevision":"312ab97e876e17341bf16f9e95397e14c2ac2720"}}}'
kubectl -n argocd patch application informed-decision --type merge \
  -p '{"operation":{"initiatedBy":{"username":"founder"},"sync":{"revision":"312ab97e876e17341bf16f9e95397e14c2ac2720","prune":false,"syncOptions":["ApplyOutOfSyncOnly=true"]}}}'
kubectl -n informed-decision rollout status deploy/informed-decision --timeout=300s
helm -n flex-auth rollback flex-auth-informed-decision-sitting 13 --wait     # access-engine's v13 release
```

This restores image `f5588f49…`, runtime ConfigMap `…6c221878d051` (v13 pin), the registry without the executor
row, and the NetworkPolicy without the sso peer. The new immutable ConfigMap `…b488b96a0953` is left in place
because prune is off. It is harmless. Any present-existing memo posed during the window stays in the store but
cannot be reviewed under v13 (`unknown_submitter`): it fails closed. Record it, and do not delete it.

## Step 6: record

Write `docs/evidence/<date>-present-existing-release/` (pins, Argo revision, Helm revision, backup sha256,
probe outputs without tokens) and append a "Present-existing release" section to `deploy/lifecycle/README.md`.
Then update INFD-WP-0010-T02.