Atom-oh · Nov 11, 2026 · Nov 11, 2026
diff --git a/docs/runbooks/AGENTS.md b/docs/runbooks/AGENTS.md
@@ -1,4 +1,4 @@
-<!-- generated-by: co-agent · source: CLAUDE.md · claude-md-sha: 82ed41aabe32 · generated-at: 2026-11-06 · DO NOT EDIT — edit CLAUDE.md then run /co-agent sync-context -->
+<!-- generated-by: co-agent · source: CLAUDE.md · claude-md-sha: 204c9c831227 · generated-at: 2026-11-11 · DO NOT EDIT — edit CLAUDE.md then run /co-agent sync-context -->
 
 > You are an external reviewer for this repo — project context below, distilled from CLAUDE.md. This file is shared verbatim by the external review panel (not a per-AI copy).
 
@@ -226,13 +226,13 @@ distinction between operational degraded data and ineligible release evidence.
 Four lanes share Lambda/Steampipe limits; contention can fail the gate. Budgets are
 admission bounds, not guaranteed completion. Runtime foundation records a 43-type,
 57.461-second collection sample and its limits; it is not full live readiness.
-Require every pinned baseline name (currently 46, source-AST checked); valid growth
+Require every pinned baseline name (currently 49, source-AST checked); valid growth
 is allowed up to 128 and every returned type needs strict proof. Both modes default
 to the enabled host only; explicit applied targets require exact enabled registration.
 Collect additionally requires measured SQL reachability with zero unreachable accounts
 and fresh account-bound EC2/CloudFront known-member proof. Explicit scope preserves null for host-only/unmeasured
 counts; CI permits host-only/null only for five source-AST-pinned SDK types, not
-46-type coverage for every member. Only Terraform onboarding
+49-type coverage for every member. Only Terraform onboarding
 preflight permits approved subsets; apply reads the restored plan, not a newer secret.
 Runtime release stays exact. Preserve isolated
 AWS CLI environments, first chronological terminal failure and stopped admission;

diff --git a/docs/runbooks/CLAUDE.md b/docs/runbooks/CLAUDE.md
@@ -280,13 +280,13 @@ distinction between operational degraded data and ineligible release evidence.
 Four lanes share Lambda/Steampipe limits; contention can fail the gate. Budgets are
 admission bounds, not guaranteed completion. Runtime foundation records a 43-type,
 57.461-second collection sample and its limits; it is not full live readiness.
-The catalog must include the pinned baseline (currently 46 names, source-AST checked);
+The catalog must include the pinned baseline (currently 49 names, source-AST checked);
 valid growth is allowed up to 128 types and every returned type needs strict proof.
 Both modes default to the enabled host only. Explicit applied targets require exact
 enabled host/member registration. Collect additionally requires measured SQL reachability
 with zero unreachable accounts and fresh account-bound EC2/CloudFront known-member evidence.
 Host-only/unmeasured counts remain null under explicit `account_reachability_scope`;
-CI permits host-only/null only for five source-AST-pinned SDK types, not 46-type
+CI permits host-only/null only for five source-AST-pinned SDK types, not 49-type
 coverage for every member. Terraform plan/apply onboarding
 alone allows approved subsets; apply derives scope from the restored saved plan.
 Runtime release never requests subset leniency. See `runtime-foundation.md#explicit-runtime-targets`.

diff --git a/docs/runbooks/graph-read-contract.md b/docs/runbooks/graph-read-contract.md
@@ -292,7 +292,7 @@ healthy evidence. Qualified empty/unproven telemetry retains the previous graph.
 Failed, retained, skipped or unattempted host infra still withholds trace collection.
 A clean confirmed-empty host publication is valid; zero nodes alone are not proof.
 Member-only gaps remain in fleet-wide counts/reasons and CLI exit 1/2. None of these
-presentation rules relaxes the strict 46-type, zero-unknown runtime release gate.
+presentation rules relaxes the strict 49-type, zero-unknown runtime release gate.
 
 | Exact diagnostic | Meaning and next check |
 |---|---|

diff --git a/docs/runbooks/runtime-foundation.md b/docs/runbooks/runtime-foundation.md
@@ -125,7 +125,7 @@ After registration, release prepare and collect require **exactly** the enabled
 and every configured target, each unique. The controller never requests onboarding
 leniency. Missing, disabled or unexpected enabled members stop before type collection.
 
-Collect retains every aggregate catalog proof (currently 46), known row/unknown-attribute
+Collect retains every aggregate catalog proof (currently 49), known row/unknown-attribute
 counts and zero unknown attributes. In target mode, SQL types must report
 `account_reachability_scope="enabled_scan_accounts"` and numeric
 `unreachable_account_count=0`. This measures the enabled, renderable DB scan scope
@@ -295,7 +295,7 @@ Group import uses a slash; membership uses comma-separated pool/group/username.
 
 Before invocation, the controller compares live `CodeSha256` with `runtime_deployment.inventory.sync_code_sha256`, derived from configured `source_code_hash`. Apply reviewed configuration to persist this expected fingerprint; provider observations cannot authorize unreviewed code. The controller also captures the Lambda revision and rechecks the hash and revision after all owned collection calls settle, before authenticated acceptance. A concurrent code/configuration change rejects that evidence, including a change restored to the same hash with a different revision.
 
-Every current catalog type (46 in this version) must have succeeded after the release marker, with known counts and zero unknown attributes. Partial, failed, stale, missing or unknown evidence blocks the release. A recent success from before the marker cannot substitute. The controller obtains the complete catalog from the verified Lambda and synchronously invokes each type through at most four concurrent collectors. It never submits `type=all` or asynchronous Event batches; existing scheduled work can still contend with its calls.
+Every current catalog type (49 in this version) must have succeeded after the release marker, with known counts and zero unknown attributes. Partial, failed, stale, missing or unknown evidence blocks the release. A recent success from before the marker cannot substitute. The controller obtains the complete catalog from the verified Lambda and synchronously invokes each type through at most four concurrent collectors. It never submits `type=all` or asynchronous Event batches; existing scheduled work can still contend with its calls.
 
 After catalog discovery, authenticated preparation verifies login, DB and host registration and samples Aurora's UTC clock. That `server_time` becomes the marker before all collection calls and remains fixed across retries. The controller calibrates later clock reads from the DB sample and local request-start timestamp, conservatively, and shifts the existing overall deadline by the same offset. It never anchors at response end or allows pre-marker data. DB-request and host-check elapsed time consume the marker window; malformed or missing clock evidence stops collection. Catalog admission allows up to 450 seconds. All type attempts and their retries share the remaining collection admission window; there is no separate fifteen-minute allowance per type. An invocation needs 450 seconds remaining for the verified function timeout of at most 420 seconds plus transport. Only confirmed throttling, busy and exact superseded outcomes retry, after ten seconds. Denied, uncertain-delivery, partial, failed, unknown and malformed outcomes cannot prove collection. All admitted workers settle before private files are cleaned.
 
@@ -453,9 +453,9 @@ milliseconds and request/response times inside the observed prepare interval, wi
 DB request elapsed time from zero through 35 seconds. The DB timestamp becomes
 `collectionStartedAt`; subsequent time is `rawNow + (DB time - request start)`.
 The existing controller deadline shifts by that same offset, preserving time remaining.
-Every current catalog type (46 at this revision) must succeed after that marker with
+Every current catalog type (49 at this revision) must succeed after that marker with
 known counts and zero unknown attributes. The returned catalog must contain every
-member of the controller's pinned baseline of 46 type names, not merely 46 arbitrary
+member of the controller's pinned baseline of 49 type names, not merely 49 arbitrary
 names. A source-AST regression ties that baseline to the checked-in `QUERIES` and
 `SDK_SYNCS` catalogs. Valid additional returned types are allowed up to 128 total;
 every returned type remains required, permitting growth without replacing a baseline member.
@@ -517,7 +517,7 @@ are added. Fixed diagnostics and private cleanup remain required on success and
 The [owner's 2026-09-14 acceptance condition](https://github.com/aws-samples/sample-awsops/pull/67#issuecomment-5663692939)
 requires all current catalog types, superseding the earlier CloudFront-only verifier
 proposal. At this revision that means at least one catalog request plus at least one request
-for each of the 46 types; catalog and per-type retries add calls. Four is the concurrency ceiling, not the total call count or
+for each of the 49 types; catalog and per-type retries add calls. Four is the concurrency ceiling, not the total call count or
 a claim of fourfold throughput. The [session contract](runtime-verifier-sessions.md#collection-effects-and-proof)
 authorizes exactly catalog or a verified catalog member, never empty/all/unregistered
 payloads or asynchronous `Event` invocation. No IAM scope is widened.

diff --git a/docs/runbooks/runtime-verifier-sessions.md b/docs/runbooks/runtime-verifier-sessions.md
@@ -182,7 +182,7 @@ symlink/public/non-regular input files, or an existing output policy file.
 Collect consumers must use `RequestResponse` on the pinned function's
 unqualified ARN, without a version or alias qualifier.
 The [owner acceptance condition dated 2026-09-14](https://github.com/aws-samples/sample-awsops/pull/67#issuecomment-5663692939)
-requires complete post-marker collection of all 46 current catalog types. The
+requires complete post-marker collection of all 49 current catalog types. The
 strict controller therefore sends exactly `{"type":"catalog"}`, followed by
 `{"type":"<catalog member>"}` for every validated returned type. This supersedes
 the earlier catalog/CloudFront-only consumer proposal; IAM scope is unchanged.
@@ -212,7 +212,7 @@ raw AWS errors. Each synchronous response must have `StatusCode=200`, no
 
 | Payload | Required result |
 | --- | --- |
-| `catalog` | Exactly `status: "catalog"` and a unique catalog containing every pinned baseline member (currently 46), with valid growth allowed up to 128 total. The source-AST test binds baseline membership to the checked-in collector. No result `type` or counts are expected. |
+| `catalog` | Exactly `status: "catalog"` and a unique catalog containing every pinned baseline member (currently 49), with valid growth allowed up to 128 total. The source-AST test binds baseline membership to the checked-in collector. No result `type` or counts are expected. |
 | Each catalog member | `status: "succeeded"`, exact requested `type`, nonnegative safe-integer `row_count`, and `unknown_attribute_count: 0` |
 
 `busy`, `failed` (including superseded), `partial`, unknown-type errors and
@@ -236,7 +236,7 @@ unreachable accounts and fresh account-bound known-member EC2/CloudFront evidenc
 `account_reachability_scope` distinguishes
 `enabled_scan_accounts` measurement from host-only and unmeasured results, whose counts stay
 null. CI permits host-only/null only for the five source-AST-pinned SDK types; aggregate
-catalog proof does not establish 46-type coverage for every member. See the [explicit target contract](runtime-foundation.md#explicit-runtime-targets).
+catalog proof does not establish 49-type coverage for every member. See the [explicit target contract](runtime-foundation.md#explicit-runtime-targets).
 Each member uses one exact `/api/deployment/member-inventory` lookup, whose bounded
 identity projection also rejects disabled, absent or ambiguous scan scope/evidence.
 Only Terraform onboarding preflight permits approved subsets, deriving apply scope

diff --git a/docs/runbooks/steampipe-quota-and-staleness.md b/docs/runbooks/steampipe-quota-and-staleness.md
@@ -132,7 +132,7 @@ observe a complete cycle under [safe tuning](#6-안전한-튜닝--safe-tuning).
 Apply only the exact reviewed plan outside active collector/runtime proof, then
 confirm service stability and the effective `steampipe_limiter_config` event.
 Preserve bucket 4, plugin/Lambda/collector concurrency 4, schedule, timeouts and IAM.
-Require all 46 baseline types and every current catalog type to complete with zero
+Require all 49 baseline types and every current catalog type to complete with zero
 unknown attributes, plus authentication/DB, model and both owned worker proofs.
 Do not remove fields, accept fallback data as complete, or disable gates to pass.
 
@@ -404,7 +404,7 @@ and unknown attributes still stop release, and every configured member still nee
 fresh known-resource evidence alongside the aggregate catalog proof.
 `collection_attempts` retains this bounded scope; null scope means no valid scope
 was reported yet. These RPC/log fields do not turn the singleton ledger into a
-per-account coverage report or assert that all 46 types cover each member.
+per-account coverage report or assert that all 49 types cover each member.
 Enabled accounts with no enabled region and `all_regions=false` are excluded by
 `_enabled_target_accounts`; the metric must not imply they were measured. In explicit
 target mode the renderer rejects such registered members, and the exact member-proof
@@ -504,6 +504,8 @@ concurrency change needs a whole-plan check showing no DNS changes.
 
 The values are safeguards, not assertions of universal AWS quotas; service, operation, account, and Region quotas differ.
 
+**ENI inventory sizing.** The `eni` type has no pushdown predicate. Every ENI is written in one Aurora round trip, and stale ENIs are deleted one per row, all inside the fixed `AURORA_RESERVE_S` (120 s). Fargate task churn enlarges the stale set on every run. At the 2026-11-11 measurement (about 470–520 ENIs per account in ap-northeast-2, 41 in us-east-1) this is far below the reserve. Watch the `eni` sync duration in the sync Lambda logs as accounts or regions grow. If one account approaches several thousand ENIs per region, batch the upserts and deletes before adding more targets. An `eni` overrun fails that type, and the runtime release gate requires every catalog type, so it blocks web releases until the type succeeds.
+
 ## 7. 롤백 / Rollback
 
 롤백은 파괴적 데이터베이스 변경 없이 이전 limiter defaults 또는 AgentCore catalog를 복원하는 방식이다.

diff --git a/scripts/AGENTS.md b/scripts/AGENTS.md
@@ -1,4 +1,4 @@
-<!-- generated-by: co-agent · source: CLAUDE.md · claude-md-sha: c60071532258 · generated-at: 2026-11-06 · DO NOT EDIT — edit CLAUDE.md then run /co-agent sync-context -->
+<!-- generated-by: co-agent · source: CLAUDE.md · claude-md-sha: 73b0f08cc5cb · generated-at: 2026-11-11 · DO NOT EDIT — edit CLAUDE.md then run /co-agent sync-context -->
 
 > You are an external reviewer for this repo — project context below, distilled from CLAUDE.md. This file is shared verbatim by Kiro, Codex, and Agy (not a per-AI copy).
 
@@ -262,7 +262,7 @@ verification. Dev releases verify exact ECS/image proof first and pass `EXPECTED
 Full releases require collect mode (including login/DB); prepare never establishes full readiness.
 Explicit activation remains separate and inactive prerequisites cannot be skipped.
 It binds dev source/account/actual role, applied runtime identity and ARM64 web digest.
-Require the pinned 46-name baseline, source-AST checked; valid growth is allowed up
+Require the pinned 49-name baseline, source-AST checked; valid growth is allowed up
 to 128 types. Every returned type
 needs post-marker succeeded evidence, known counts and
 zero unknowns, with at most four concurrent in-flight synchronous owned calls. Prepare obtains authenticated
@@ -274,7 +274,7 @@ host/member registration, measured SQL reachability over enabled scan accounts w
 unreachable accounts and fresh account-bound EC2/CloudFront member evidence from one
 exact minimal `/api/deployment/member-inventory` lookup. Explicit scope preserves null for
 host-only/unmeasured counts; CI allows host-only/null only for five source-AST-pinned
-SDK types, not 46-type coverage for every member. Only Terraform onboarding preflight permits approved subsets;
+SDK types, not 49-type coverage for every member. Only Terraform onboarding preflight permits approved subsets;
 apply reads the restored saved plan, never a newer secret. Runtime release stays exact.
 AWS CLI children use an explicit
 credential/settings allowlist, pinned path, disabled config/credential files/metadata

diff --git a/scripts/CLAUDE.md b/scripts/CLAUDE.md
@@ -405,9 +405,9 @@ full readiness (including login/DB); prepare only verifies
 authenticated login/DB/host registration. Explicit runtime/readiness activation remains
 separate, and inactive prerequisites cannot be skipped.
 The controller verifies dev source/account/role, applied runtime metadata and ARM64
-web digest, requires every pinned baseline member (currently 46, source-AST checked)
+web digest, requires every pinned baseline member (currently 49, source-AST checked)
 and permits valid growth up to 128 types, then drives every returned
-type (currently 46) with at most four
+type (currently 49) with at most four
 concurrent in-flight synchronous invocations. It requires succeeded results, known counts and zero unknowns.
 It samples the authenticated DB clock before collecting, anchors calibration at request
 start, shifts the existing deadline by the same offset, and retains strict post-marker
@@ -418,7 +418,7 @@ require the exact enabled host/member registry before type calls, measured SQL r
 with zero unreachable accounts, and fresh account-bound known EC2/CloudFront member rows.
 `account_reachability_scope` distinguishes `enabled_scan_accounts` measurement from host-only
 and unmeasured results; the latter counts stay null. CI permits host-only/null only for
-the five source-AST-pinned SDK types. The aggregate 46-type proof does not establish 46-type coverage for every member.
+the five source-AST-pinned SDK types. The aggregate 49-type proof does not establish 49-type coverage for every member.
 Member evidence uses one exact, bounded `/api/deployment/member-inventory` response,
 never a scan of full inventory pages.
 Only Terraform plan/apply onboarding preflight permits approved subsets; apply reads

diff --git a/scripts/v2/ci/runtime-release.mjs b/scripts/v2/ci/runtime-release.mjs
@@ -26,7 +26,8 @@ export const REQUIRED_CATALOG_TYPES = Object.freeze([
   'transit_gateway_attachment', 'vpc_peering_connection', 'vpc_endpoint',
   'elasticache_replication_group', 'iam_policy', 'neptune_cluster', 'msk', 'waf',
   'waf_rule_group', 'waf_ip_set', 'cloudwatch_alarm', 'cloudtrail', 'apigatewayv2_api',
-  'apigatewayv2_integration', 'apigatewayv2_route', 'ebs_snapshot', 's3',
+  'apigatewayv2_integration', 'apigatewayv2_route', 'ebs_snapshot', 'eni', 'rds_cluster',
+  'rds_proxy', 's3',
   'opensearch_serverless', 'cloudfront_vpc_origin', 'alb_listener_rule', 's3_public_access',
 ]);
 export const MIN_CATALOG_TYPES = REQUIRED_CATALOG_TYPES.length;