nocturne-memory 0.1.12__tar.gz

nocturne_memory-0.1.12/.gitignore

@@ -0,0 +1,9 @@
+.env
+.pytest_cache/
+.ruff_cache/
+.venv/
+__pycache__/
+*.py[cod]
+*.egg-info/
+dist/
+/verification/*/impacted.json

nocturne_memory-0.1.12/Dockerfile

@@ -0,0 +1,28 @@
+FROM python:3.12.13-slim@sha256:57cd7c3a7a273101a6485ba99423ee568157882804b1124b4dd04266317710de AS base
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1
+
+FROM base AS builder
+
+WORKDIR /build
+
+COPY pyproject.toml ./
+COPY README.md ./
+COPY Dockerfile ./
+COPY src ./src
+COPY infra ./infra
+
+RUN pip wheel --no-cache-dir --no-deps --wheel-dir /wheels .
+
+FROM base AS runtime
+
+WORKDIR /app
+
+COPY --from=builder /wheels /wheels
+
+RUN pip install --no-cache-dir /wheels/nocturne_memory-*.whl && rm -rf /wheels
+
+EXPOSE 8000
+
+CMD ["uvicorn", "spine.main:create_app", "--factory", "--host", "0.0.0.0", "--port", "8000"]

nocturne_memory-0.1.12/PKG-INFO

@@ -0,0 +1,46 @@
+Metadata-Version: 2.5
+Name: nocturne-memory
+Version: 0.1.12
+Summary: NOCTURNE Memory Palace service
+Project-URL: Homepage, https://github.com/Nate0-1999/nocturne
+Project-URL: Source, https://github.com/Nate0-1999/nocturne-memory
+Requires-Python: <3.13,>=3.12
+Requires-Dist: alembic<2,>=1.13
+Requires-Dist: asyncpg<1,>=0.29
+Requires-Dist: fastapi<1,>=0.115
+Requires-Dist: httpx<1,>=0.27
+Requires-Dist: pgvector<1,>=0.3
+Requires-Dist: pydantic-settings<3,>=2
+Requires-Dist: pydantic<3,>=2
+Requires-Dist: sqlalchemy[asyncio]<3,>=2
+Requires-Dist: tiktoken<1,>=0.9
+Requires-Dist: uvicorn[standard]<1,>=0.30
+Provides-Extra: dev
+Requires-Dist: pytest-asyncio<1,>=0.24; extra == 'dev'
+Requires-Dist: pytest<9,>=8; extra == 'dev'
+Requires-Dist: ruff<1,>=0.9; extra == 'dev'
+Requires-Dist: testcontainers[postgres]<5,>=4.8; extra == 'dev'
+Description-Content-Type: text/markdown
+
+# nocturne-memory
+
+`nocturne-memory` is NOCTURNE's Memory Palace service and migration package. It
+is released in lockstep as a dependency of the public `nocturne-harness`
+distribution.
+
+The service owns the extraction consent boundary at `/v1/extractions` and
+`/v1/approval-queue`. Pending units use the internal `candidate` lifecycle
+state and are excluded from ordinary memory listing, search, and injection.
+Queue decisions are append-only; approval enacts the proposed verdict through
+revisioned head changes and typed lineage edges, while denial revision-
+tombstones the candidate as rejected.
+
+Injection treats memory room as a versioned share of the active model context
+window. The share is a ceiling, not a quota: unused room remains available to
+conversation, while pinned memories always inject and report any overflow in
+the prepare allocation. Share and threshold join the ordinary replay learner
+after 100 authentic, hygiene-filtered owner dispositions; every learned
+generation remains proposed until the owner activates it.
+
+New users should start at the [NOCTURNE quickstart](https://github.com/Nate0-1999/nocturne)
+and install `nocturne-harness`; they do not need to install this package separately.

nocturne_memory-0.1.12/README.md

@@ -0,0 +1,22 @@
+# nocturne-memory
+
+`nocturne-memory` is NOCTURNE's Memory Palace service and migration package. It
+is released in lockstep as a dependency of the public `nocturne-harness`
+distribution.
+
+The service owns the extraction consent boundary at `/v1/extractions` and
+`/v1/approval-queue`. Pending units use the internal `candidate` lifecycle
+state and are excluded from ordinary memory listing, search, and injection.
+Queue decisions are append-only; approval enacts the proposed verdict through
+revisioned head changes and typed lineage edges, while denial revision-
+tombstones the candidate as rejected.
+
+Injection treats memory room as a versioned share of the active model context
+window. The share is a ceiling, not a quota: unused room remains available to
+conversation, while pinned memories always inject and report any overflow in
+the prepare allocation. Share and threshold join the ordinary replay learner
+after 100 authentic, hygiene-filtered owner dispositions; every learned
+generation remains proposed until the owner activates it.
+
+New users should start at the [NOCTURNE quickstart](https://github.com/Nate0-1999/nocturne)
+and install `nocturne-harness`; they do not need to install this package separately.

nocturne_memory-0.1.12/api_contract_fingerprints.json

@@ -0,0 +1,14 @@
+{
+  "0.1.0": "ebc263f97e4b357e3370d2b5a188f752bf67b1743d6cde9bd68acff853703e33",
+  "0.1.1": "ebc263f97e4b357e3370d2b5a188f752bf67b1743d6cde9bd68acff853703e33",
+  "0.1.10": "503eaed7b386b0a96ab56a444081f1a37a28f113cbc6e6617e434af82765794d",
+  "0.1.11": "304172efb12e804494edb7cccf11f2d1d00ef69d0e204fa659595dddf10dc20c",
+  "0.1.2": "ec1b87185670ac86377b389ac67c536fb295dac029e4e97c4ffcccd3017cee5b",
+  "0.1.3": "4fc4d8920ce54ea0766c933719429bf14e7372c1a5ab6f80a5b2d5d42adf699b",
+  "0.1.4": "cb7499803498e8097529f3812b9bfeb094f100ff6761b29e461c56b1674ae38d",
+  "0.1.5": "3867455375aabcf861c87af0432d0c27f710cdb13aac199e86f2227a9a16777b",
+  "0.1.6": "48d16b6a328e61dd31a9f84691b4a01fc4107b13adc99955de18dec22278dce2",
+  "0.1.7": "b8ff406c1f57d0bfc2a9a175940da48dd4d258b11d4b0695ecba44e8aec910b3",
+  "0.1.8": "d182caa605cd1430c7fba0012ade031cb78ab5308043d25e18aaf552500e182a",
+  "0.1.9": "662ac885d8a5f39f10464d2f5db545aed435fa039220cf5d65c22228e653ba94"
+}

nocturne_memory-0.1.12/infra/billing-breaker/README.md

@@ -0,0 +1,545 @@
+# D2 billing circuit breaker
+
+> STATUS: DEPLOYED AND ARMED in `n8-memory-palace` on 2026-07-21. The
+> `billing-breaker` topic, private no-retry function, isolated identities,
+> detach-role binding, and budget wiring are live and least-privilege
+> verified. To change or remove it, use the cleanup procedure below, then
+> re-run `deploy.sh --apply` (it is first-deploy-only and refuses to run
+> while these resources exist). The preflight was hardened at deploy time to
+> run against a default-posture GCP project (recognizes Google default
+> identities: Compute, App Engine `@appspot`, Container Registry, and
+> `gcp-sa-*` agents); see DECISIONS 017–018.
+
+This package is an armed, one-way cost control for `n8-memory-palace`. A valid
+budget message with `costAmount >= budgetAmount` calls Cloud Billing
+`updateBillingInfo` with an empty `billingAccountName`. That deliberately
+detaches the project's billing account and can stop or irretrievably delete
+Cloud Run, Cloud SQL, and every other billable project service. There is no
+simulation switch in the deployed function.
+
+Agents may build and test this directory. **Only a human may run `deploy.sh
+--apply`, publish a live message, alter IAM, or execute recovery commands.**
+
+## What this does
+
+- accepts only Pub/Sub schema `1.0` messages whose `billingAccountId` and
+  `budgetId` match deploy-time settings;
+- keeps the destructive target fixed in source as `projects/n8-memory-palace`;
+- compares JSON numbers as decimals and detaches at equality or overage;
+- emits one-line structured JSON decisions before and after the action;
+- acknowledges malformed, foreign, and below-budget messages without creating
+  a Cloud Billing client;
+- surfaces Cloud Billing failures, while the deployed trigger explicitly does
+  not retry failed events; the next periodic budget status is the next attempt;
+- makes duplicate delivery state-idempotent by repeatedly requesting the same
+  empty billing-account assignment, without a broader status-read permission.
+
+Google documents the [budget notification schema, periodic delivery, and
+at-least-once semantics](https://docs.cloud.google.com/billing/docs/how-to/budgets-programmatic-notifications),
+the [empty-account detach operation](https://docs.cloud.google.com/billing/docs/reference/rest/v1/projects/updateBillingInfo),
+the [disable-billing pattern and blast radius](https://docs.cloud.google.com/billing/docs/how-to/disable-billing-with-notifications),
+and the [function retry hazard](https://docs.cloud.google.com/run/docs/tips/function-retries).
+Retries are disabled because a permanent IAM/configuration failure could
+otherwise keep an old over-budget event alive and detach again after recovery.
+
+This is not a hard or real-time $100 cap. Budget data is estimated and delayed,
+notifications arrive only several times per day, delivery can be duplicated or
+out of order, and already-incurred but unreported charges remain payable.
+
+## Human preflight
+
+Stop if any item is uncertain:
+
+1. Put the Cloud SQL backup and D1 redeploy material **outside
+   `n8-memory-palace`**. A backup in the project can become unavailable with the
+   project it is meant to restore. Google warns that disabling billing can
+   remove resources and recovery is not guaranteed.
+2. In Cloud Billing Reports, verify the current-period **actual** cost is below
+   USD 100 immediately before arming. If it is already at or over USD 100,
+   connecting the budget is an intentional near-term outage.
+3. Confirm the existing budget has `ownershipScope: BILLING_ACCOUNT`, is
+   recurring monthly, specified USD 100.00, filtered to only the numeric project
+   resource for `n8-memory-palace`, and has no service, label, subaccount,
+   resource-ancestor, custom-period, or credit filter. Its Pub/Sub notification
+   topic must be disconnected.
+4. Confirm the project-to-billing-account link is unlocked and the account has
+   no active or pending commitment that prevents detachment.
+5. Audit inherited/custom access on both the project and the named billing
+   account for `billing.budgets.update`, `billing.resourcebudgets.write`,
+   `billing.accounts.setIamPolicy`, Pub/Sub topic/subscription update or IAM,
+   Eventarc trigger update or IAM, topic subscription attachment/detachment,
+   `pubsub.topics.publish`, `run.routes.invoke`, billing association changes,
+   `iam.serviceAccounts.actAs`, token/key minting, and IAM-policy mutation.
+   Pub/Sub attributes are not signatures, and a principal that can retarget or
+   lower the exact budget can make Google's publisher emit an authentic outage
+   event. The budget must use `ownershipScope: BILLING_ACCOUNT`, making project
+   principals read-only even when they otherwise have resource-budget write
+   permission. For every direct project and billing-account binding, the script
+   reads
+   the role's current `includedPermissions`; only the active human may directly
+   control the budget or billing-account IAM (matched case-insensitively on the
+   email, since Google account emails are case-insensitive). It rejects other
+   dangerous project permissions except the exact armed runtime binding and
+   Google-managed, project-number-pinned service agents — the default Cloud
+   Build, cloudservices, and Compute Engine accounts, a curated set of
+   Google-owned service-agent domains, and Google's reserved `gcp-sa-*` per-
+   service agents. These identities are created and controlled by Google for
+   this project and are not freely mintable, so a default-posture project is
+   deployable; a service account minted in any other project (its own
+   `<project-id>` SA domain) stays audited. It separately requires
+   empty direct policies and no user-managed keys on all fresh D2 identities,
+   only Google's budget publisher on the topic, and only the trigger identity
+   on the service. The armed path must have one target-project Eventarc
+   subscription that exactly equals the function-owned Eventarc trigger's
+   output-only transport, a healthy exact destination, no direct subscription
+   policy, and no topic or subscription Single Message Transforms because those
+   can rewrite both payload data and attributes. Inherited roles, group
+   membership, and humans who can command
+   a trusted service agent are not fully resolved by those direct policies and
+   remain a manual destructive-admin boundary.
+6. Use a human identity, not service-account credentials. Remove all effective
+   gcloud impersonation, access-token-file, and credential-file overrides; the
+   script refuses them before trusting `gcloud auth list`. The simplest
+   auditable gate is the sole direct Project Owner on `n8-memory-palace` plus
+   Billing Account Administrator on the named account. Any other direct
+   principal with a dangerous current permission makes the script stop. The
+   named `roles/billing.projectManager` binding must be absent before arming;
+   D2 reserves that exact role for its runtime identity. Recovery still needs
+   the permissions documented by Google: Project Billing Manager + Browser +
+   Service Usage Viewer on the project and Billing Account User + Viewer on the
+   account, obtained only after the breaker is torn down if Owner is not used.
+
+The [billing-link permissions](https://docs.cloud.google.com/billing/docs/how-to/modify-project#required_permissions)
+matter during recovery: Billing Account Costs Manager alone cannot reattach a
+project.
+
+Identify the exact existing budget; never select it by display name:
+
+```sh
+export BILLING_ACCOUNT_ID=000000-000000-000000
+gcloud billing budgets list --billing-account="${BILLING_ACCOUNT_ID}"
+export BUDGET_RESOURCE="billingAccounts/${BILLING_ACCOUNT_ID}/budgets/BUDGET_ID"
+gcloud billing budgets describe "${BUDGET_RESOURCE}"
+```
+
+## Review, then deploy
+
+The dry run is local-only: it validates its two arguments and prints the plan
+without looking for or invoking `gcloud`.
+
+```sh
+cd infra/billing-breaker
+export BILLING_ACCOUNT_ID=000000-000000-000000
+export BUDGET_RESOURCE="billingAccounts/${BILLING_ACCOUNT_ID}/budgets/BUDGET_ID"
+./deploy.sh --dry-run
+```
+
+`--apply` is deliberately first-deploy-only. The function, Cloud Run service,
+topic, every Eventarc trigger that references a D2 name, and all three service
+accounts must be absent. A failed deployment or a rearm must go through the
+cleanup procedure instead of silently adopting unknown identities, keys,
+triggers, or queued messages.
+
+After reviewing the plan and completing the preflight, a human can enter the
+interactive gate:
+
+```sh
+export CONFIRM_D2_PROJECT=n8-memory-palace
+./deploy.sh --apply
+```
+
+The script repeats the project, budget, account, destructive action, and
+below-USD-100 assertion in a typed confirmation before its first mutation. It
+rejects ambient credential overrides, resolves every direct role to its current
+permissions, and uses successful list responses for its fresh-resource checks,
+rather than treating any failed `describe` as absence. It then:
+
+1. validates the enabled billing link and exact disconnected budget;
+2. creates a fresh `billing-breaker` topic;
+3. separates the runtime, Eventarc trigger, and build identities;
+4. gives the build identity only Google's documented build roles and removes
+   all three immediately when the build completes;
+5. deploys a private Python 3.12 second-generation function with retries off;
+6. gives the trigger identity Invoker only on this Cloud Run service and gives
+   only Google's budget publisher direct topic Publisher access;
+7. validates the complete inert topology;
+8. requires the named Project Billing Manager role to be absent, then gives its
+   sole unconditional membership to the runtime identity on this project
+   immediately before connecting the budget last; and
+9. asserts the complete armed state. Any later error enters a rollback that
+   revokes applicable runtime/build roles and reads the project policy back.
+   If absence cannot be proved, the script exits with a `CRITICAL` warning and
+   the operator must assume detach authority remains active.
+
+The runtime receives no billing-account, Artifact Registry, build, Eventarc,
+or project-wide Invoker role. The trigger cannot detach billing, and the build
+identity has no standing project role after deployment. Google documents the
+[custom build identity roles](https://docs.cloud.google.com/functions/docs/building#secure_your_build_with_a_custom_service_account).
+
+## Synthetic-message drill
+
+The live drill is intentionally below budget. Equality and overage are proven
+with fake-client unit tests; publishing either value to the armed topic would
+perform the outage this breaker exists to cause. Do not add a new Publisher
+binding merely to run the drill; use an already-trusted Project Owner or skip
+the live drill.
+
+```sh
+export BUDGET_ID="${BUDGET_RESOURCE##*/}"
+if ! drill_message_id="$(
+  gcloud pubsub topics publish billing-breaker \
+    --project=n8-memory-palace \
+    --attribute="billingAccountId=${BILLING_ACCOUNT_ID},budgetId=${BUDGET_ID},schemaVersion=1.0" \
+    --message='{"budgetDisplayName":"D2 safe drill","costAmount":99.99,"costIntervalStart":"2026-07-01T00:00:00Z","budgetAmount":100.00,"budgetAmountType":"SPECIFIED_AMOUNT","currencyCode":"USD"}' \
+    --format='value(messageIds)'
+)"; then
+  echo "publish failed; drill did not pass" >&2
+  exit 1
+fi
+if [[ -z "${drill_message_id}" || "${drill_message_id}" == *$'\n'* ]]; then
+  echo "publish did not return exactly one message ID" >&2
+  exit 1
+fi
+log_filter="resource.type=\"cloud_run_revision\" AND resource.labels.service_name=\"billing-breaker\" AND jsonPayload.component=\"billing-breaker\" AND jsonPayload.action=\"below_budget\" AND jsonPayload.message_id=\"${drill_message_id}\""
+if ! drill_log="$(
+  gcloud logging read "${log_filter}" \
+    --project=n8-memory-palace --limit=1 --format=json
+)"; then
+  echo "log lookup failed; drill did not pass" >&2
+  exit 1
+fi
+DRILL_LOG="${drill_log}" python3 - <<'PY'
+import json
+import os
+
+if not json.loads(os.environ["DRILL_LOG"]):
+    raise SystemExit("no below_budget decision matched the published message ID")
+PY
+if ! billing_json="$(
+  gcloud billing projects describe n8-memory-palace --format=json
+)"; then
+  echo "billing lookup failed; drill did not pass" >&2
+  exit 1
+fi
+if ! printf '%s' "${billing_json}" | python3 deployment_checks.py billing \
+  --billing-account-id "${BILLING_ACCOUNT_ID}"; then
+  echo "billing validator failed; drill did not pass" >&2
+  exit 1
+fi
+```
+
+Pass means the exact published message ID has a `below_budget` JSON decision
+and the machine-checked billing link remains enabled on the expected account.
+Eventarc can take up to two minutes to propagate; an empty log result is a
+failed/incomplete drill, so wait and repeat only the log and billing checks.
+Do not republish merely because the log is delayed. Do not publish a message at
+or above 100 to the live topic as a routine test. At-least-once delivery can
+duplicate even a successful publish.
+
+## Recovery and billing reattach
+
+The order is a safety invariant. Do not restore the runtime role while an old
+trigger, subscription, or topic exists, and do not reattach billing until the
+breaker has neither authority nor a queue.
+
+Set the exact resources first and run every block in the same Bash session.
+Every absence proof below comes from a successful list request plus the same
+exact-value validator used by deployment. Every validator is explicitly
+guarded; permission, network, and parsing errors stop recovery.
+
+```sh
+: "${BUDGET_RESOURCE:?set the exact existing budget resource}"
+: "${BILLING_ACCOUNT_ID:?set the billing account ID}"
+```
+
+1. In Cloud Billing **Budgets & alerts**, edit `BUDGET_RESOURCE`, disconnect
+   its Pub/Sub topic, save, and machine-check that the field is empty:
+
+   ```sh
+   if ! budget_topic="$(
+     gcloud billing budgets describe "${BUDGET_RESOURCE}" \
+       --format='value(notificationsRule.pubsubTopic)'
+   )"; then
+     echo "budget lookup failed; disconnection is not proved" >&2
+     exit 1
+   fi
+   if [[ -n "${budget_topic}" ]]; then
+     printf 'budget is still connected to %s\n' "${budget_topic}" >&2
+     exit 1
+   fi
+   ```
+
+2. Revoke detach authority and verify the named role is completely absent:
+
+   ```sh
+   gcloud projects remove-iam-policy-binding n8-memory-palace \
+     --member="serviceAccount:billing-breaker-runtime@n8-memory-palace.iam.gserviceaccount.com" \
+     --role=roles/billing.projectManager || true
+   if ! project_policy="$(
+     gcloud projects get-iam-policy n8-memory-palace --format=json
+   )"; then
+     echo "project policy lookup failed; detach-role removal is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${project_policy}" | \
+     python3 deployment_checks.py exact-project-role \
+       --role=roles/billing.projectManager --state=absent; then
+     echo "detach-role validator failed; removal is not proved" >&2
+     exit 1
+   fi
+   ```
+
+   The removal command is allowed to report an already-absent binding, but the
+   policy read and exact empty-role validator must both succeed.
+
+3. Capture every subscription attached to the dedicated topic from the
+   topic-side Pub/Sub list, including subscriptions owned by another project,
+   and independently capture every matching Eventarc trigger. Then delete the
+   function, triggers, Cloud Run service, and all captured subscriptions. This
+   sweep includes an orphan left by a partial Eventarc teardown. If the human
+   cannot delete a cross-project subscription, detach it using topic authority;
+   detached messages are deleted and the subscription cannot be reattached.
+   The current `gcloud functions delete` command does not take `--gen2`.
+   Deleting a Pub/Sub topic alone does not delete its subscriptions or retained
+   backlog, so topic-side absence is proved before topic deletion. Google
+   documents [topic deletion](https://docs.cloud.google.com/pubsub/docs/delete-topic)
+   and [subscription detachment](https://docs.cloud.google.com/pubsub/docs/detach-subscriptions).
+
+   ```sh
+   if ! subscription_before_json="$(
+     gcloud pubsub topics list-subscriptions billing-breaker \
+       --project=n8-memory-palace --format=json
+   )"; then
+     echo "subscription lookup failed; topic-wide queue capture is not proved" >&2
+     exit 1
+   fi
+   if ! subscription_rows="$(
+     printf '%s' "${subscription_before_json}" | \
+       python3 deployment_checks.py topic-subscriptions --state=list
+   )"; then
+     echo "topic-wide queue parsing failed; queue identity is not proved" >&2
+     exit 1
+   fi
+   if ! eventarc_before_json="$(
+     gcloud eventarc triggers list \
+       --location=us-central1 --project=n8-memory-palace --format=json
+   )"; then
+     echo "Eventarc lookup failed; transport capture is not proved" >&2
+     exit 1
+   fi
+   if ! trigger_rows="$(EVENTARC_JSON="${eventarc_before_json}" python3 - <<'PY'
+import json
+import os
+
+topic = "projects/n8-memory-palace/topics/billing-breaker"
+triggers = json.loads(os.environ["EVENTARC_JSON"])
+if not isinstance(triggers, list):
+    raise SystemExit("Eventarc list response is not an array")
+for trigger in triggers:
+    pubsub = trigger.get("transport", {}).get("pubsub", {})
+    if pubsub.get("topic") != topic:
+        continue
+    name = trigger.get("name")
+    if not isinstance(name, str) or not name.startswith(
+        "projects/n8-memory-palace/locations/us-central1/triggers/"
+    ):
+        raise SystemExit("matching Eventarc trigger name is malformed")
+    print(name)
+PY
+   )"; then
+     echo "Eventarc transport parsing failed; trigger identity is not proved" >&2
+     exit 1
+   fi
+   gcloud functions delete billing-breaker \
+     --region=us-central1 --project=n8-memory-palace --quiet || true
+   while IFS= read -r trigger_resource; do
+     [[ -z "${trigger_resource}" ]] && continue
+     gcloud eventarc triggers delete "${trigger_resource##*/}" \
+       --location=us-central1 --project=n8-memory-palace --quiet || true
+   done <<<"${trigger_rows}"
+   gcloud run services delete billing-breaker \
+     --region=us-central1 --project=n8-memory-palace --quiet || true
+
+   if ! function_list="$(
+     gcloud functions list --v2 --regions=us-central1 \
+       --project=n8-memory-palace --format=json
+   )"; then
+     echo "function list failed; deletion is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${function_list}" | python3 deployment_checks.py absent \
+     --field=name \
+     --value=projects/n8-memory-palace/locations/us-central1/functions/billing-breaker; then
+     echo "function validator failed; deletion is not proved" >&2
+     exit 1
+   fi
+   if ! run_service_list="$(
+     gcloud run services list --region=us-central1 \
+       --project=n8-memory-palace --format=json
+   )"; then
+     echo "Cloud Run list failed; deletion is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${run_service_list}" | python3 deployment_checks.py absent \
+     --field=metadata.name --value=billing-breaker; then
+     echo "Cloud Run validator failed; deletion is not proved" >&2
+     exit 1
+   fi
+   if ! eventarc_after_json="$(
+     gcloud eventarc triggers list \
+       --location=us-central1 --project=n8-memory-palace --format=json
+   )"; then
+     echo "Eventarc lookup failed; trigger removal is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${eventarc_after_json}" | \
+     python3 deployment_checks.py eventarc-isolation \
+       --topic-resource=projects/n8-memory-palace/topics/billing-breaker \
+       --function-resource=projects/n8-memory-palace/locations/us-central1/functions/billing-breaker \
+       --run-service-name=billing-breaker \
+       --run-service-resource=projects/n8-memory-palace/locations/us-central1/services/billing-breaker; then
+     echo "Eventarc validator failed; trigger removal is not proved" >&2
+     exit 1
+   fi
+   while IFS= read -r subscription_resource; do
+     [[ -z "${subscription_resource}" ]] && continue
+     if ! gcloud pubsub subscriptions delete "${subscription_resource}" --quiet; then
+       gcloud pubsub topics detach-subscription "${subscription_resource}" \
+         --quiet || true
+     fi
+   done <<<"${subscription_rows}"
+   if ! subscription_after_json="$(
+     gcloud pubsub topics list-subscriptions billing-breaker \
+       --project=n8-memory-palace --format=json
+   )"; then
+     echo "topic-side subscription list failed; queue removal is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${subscription_after_json}" | \
+     python3 deployment_checks.py topic-subscriptions --state=empty; then
+     echo "a topic-attached subscription remains; queue removal is not proved" >&2
+     exit 1
+   fi
+   ```
+
+   Stop if any successful list still contains the function, service, matching
+   trigger, or topic-attached subscription. Ignored delete/detach status supports
+   a partially completed cleanup; it never substitutes for the later list proof.
+
+4. Delete the dedicated topic to destroy the old delivery path. Remove any
+   temporary build roles left by a failed apply, verify no D2 project binding
+   remains, then delete all three identities. Successful lists must prove every
+   exact resource absent:
+
+   ```sh
+   gcloud pubsub topics delete billing-breaker \
+     --project=n8-memory-palace --quiet || true
+   for role in roles/artifactregistry.writer roles/logging.logWriter roles/storage.objectViewer; do
+     gcloud projects remove-iam-policy-binding n8-memory-palace \
+       --member="serviceAccount:billing-breaker-build@n8-memory-palace.iam.gserviceaccount.com" \
+       --role="${role}" || true
+   done
+   if ! project_policy="$(
+     gcloud projects get-iam-policy n8-memory-palace --format=json
+   )"; then
+     echo "project policy lookup failed; D2 IAM cleanup is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${project_policy}" | \
+     python3 deployment_checks.py exact-project-role \
+       --role=roles/billing.projectManager --state=absent; then
+     echo "detach-role validator failed; D2 IAM cleanup is not proved" >&2
+     exit 1
+   fi
+   for role in roles/artifactregistry.writer roles/logging.logWriter roles/storage.objectViewer; do
+     if ! printf '%s' "${project_policy}" | \
+       python3 deployment_checks.py project-role \
+         --role="${role}" \
+         --member="serviceAccount:billing-breaker-build@n8-memory-palace.iam.gserviceaccount.com" \
+         --state=absent; then
+       echo "build-role validator failed; D2 IAM cleanup is not proved" >&2
+       exit 1
+     fi
+   done
+   gcloud iam service-accounts delete \
+     billing-breaker-runtime@n8-memory-palace.iam.gserviceaccount.com \
+     --project=n8-memory-palace --quiet || true
+   gcloud iam service-accounts delete \
+     billing-breaker-trigger@n8-memory-palace.iam.gserviceaccount.com \
+     --project=n8-memory-palace --quiet || true
+   gcloud iam service-accounts delete \
+     billing-breaker-build@n8-memory-palace.iam.gserviceaccount.com \
+     --project=n8-memory-palace --quiet || true
+   if ! topic_list="$(
+     gcloud pubsub topics list --project=n8-memory-palace --format=json
+   )"; then
+     echo "topic list failed; deletion is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${topic_list}" | python3 deployment_checks.py absent \
+     --field=name \
+     --value=projects/n8-memory-palace/topics/billing-breaker; then
+     echo "topic validator failed; deletion is not proved" >&2
+     exit 1
+   fi
+   if ! service_account_list="$(
+     gcloud iam service-accounts list \
+       --project=n8-memory-palace --format=json
+   )"; then
+     echo "service-account list failed; deletion is not proved" >&2
+     exit 1
+   fi
+   for service_account in billing-breaker-runtime billing-breaker-trigger billing-breaker-build; do
+     if ! printf '%s' "${service_account_list}" | \
+       python3 deployment_checks.py absent \
+         --field=email \
+         --value="${service_account}@n8-memory-palace.iam.gserviceaccount.com"; then
+       echo "service-account validator failed; deletion is not proved" >&2
+       exit 1
+     fi
+   done
+   ```
+
+   Stop before identity deletion if any exact IAM absence check fails.
+
+5. Only after steps 1–4 pass, reattach the known open billing account and
+   verify both fields:
+
+   ```sh
+   export BILLING_ACCOUNT_ID=000000-000000-000000
+   gcloud billing projects link n8-memory-palace \
+     --billing-account="${BILLING_ACCOUNT_ID}"
+   if ! billing_json="$(
+     gcloud billing projects describe n8-memory-palace --format=json
+   )"; then
+     echo "billing lookup failed; reattachment is not proved" >&2
+     exit 1
+   fi
+   if ! printf '%s' "${billing_json}" | python3 deployment_checks.py billing \
+     --billing-account-id="${BILLING_ACCOUNT_ID}"; then
+     echo "billing validator failed; reattachment is not proved" >&2
+     exit 1
+   fi
+   ```
+
+   The validator requires `billingEnabled: true` and exactly
+   `billingAccounts/${BILLING_ACCOUNT_ID}`.
+
+6. Inspect Cloud SQL and Cloud Run, restore/redeploy D1 where necessary, and
+   repeat its cloud round trip. Reattaching billing does not guarantee every
+   resource resumes automatically.
+7. Rearm only when the current budget period is below its limit (or a new
+   period has begun), every old D2 resource is absent, and recovery is verified.
+   Run the full preflight and `deploy.sh --apply` again; it will create fresh
+   identities and a fresh queue.
+
+If `deploy.sh --apply` fails, its trap attempts to revoke applicable detach and
+temporary build roles and then proves their absence from a fresh project-policy
+read. A `CRITICAL` message or exit 99 means the proof failed: assume detach
+authority remains active and disconnect the budget immediately. The script
+deliberately leaves created resources for inspection. Verify billing is still
+enabled, disconnect the budget if it was wired, then perform steps 1–4 before
+any retry. Never work around the script by reusing those resources.
+
+Revisit this package when Google Spend Caps becomes generally available and
+covers the project's complete Cloud Run plus Cloud SQL spend surface.