Query Manifest Authority Isolation Plan

Goal: Make served Query start and operate without catalog DynamoDB access, while keeping Lance physical manifest metadata directly readable from a separate least-privilege table.

Architecture: Split the current all-in-one CLI Context into a full metadata/admin context and a minimal QueryContext. A pure validated cloud storage plan names independent registry and manifest tables before I/O. The metadata context opens both; Query opens only the manifest store and never provisions it. Local Query constructs local Lance plus its managed stage without opening Rocks catalog state.

Tech Stack: Rust, Tokio, DynamoDB-backed MetaStore, Lance external manifests, clap CLI, Kubernetes YAML, agent-spec, jj.

Task 1: Freeze the cloud authority plan

Files:

• Modify: crates/lake-cli/src/commands/mod.rs
• Test: crates/lake-cli/src/commands/mod.rs

1. Add failing selectors cloud_manifest_table_alias_fails_before_connect and cloud_storage_wiring_separates_registry_and_manifest_authority against a pure cloud table/storage plan.
2. Run both selectors and require real failures because no manifest-specific table or Query/metadata authority plan exists.
3. Implement bounded non-empty table-name validation, default names, alias rejection, and explicit metadata-vs-Query connection/provisioning plans.
4. Run the focused tests, fmt, git diff --check, and strict CLI clippy.

Task 2: Split served Query startup context

Files:

• Modify: crates/lake-cli/src/main.rs
• Modify: crates/lake-cli/src/commands/mod.rs
• Modify: crates/lake-cli/src/commands/serve.rs
• Modify: crates/lake-engine-lance/src/lib.rs
• Modify: crates/lake-engine-lance/src/manifest_store.rs
• Test: crates/lake-cli/src/commands/mod.rs
• Test: crates/lake-cli/src/main.rs

1. Add failing query_context_has_no_catalog_authority: local Query startup must not create the Rocks meta directory or expose registry/Metasrv state.
2. Introduce QueryContext with only engine and managed stage. Dispatch Query before the full Context::open path and adapt serve/async configuration to the minimal context.
3. In cloud mode connect Query only to the manifest table and open the pre-provisioned pair to load its v2 authority marker; never call any ensure/create API. Use a read-only manifest adapter that rejects write/delete and legacy latest-pointer migration before KV mutation. Keep metadata/admin context opening both distinct pairs.
4. Run focused CLI tests plus existing query wiring, selftest, and cloud wiring tests; then strict clippy.

Task 3: Update deployment and operator contract

Files:

• Modify: deploy/kubernetes/lake.yaml
• Modify: crates/lake-cli/tests/kubernetes_manifests.rs
• Modify: README.md
• Modify: docs/architecture.md
• Modify: docs/guides/kubernetes.md
• Modify: crates/lake-cli/AGENT.md

1. Add LAKE_MANIFEST_DYNAMODB_TABLE to the reference ConfigMap and make the Kubernetes selector require it.
2. Document four authorities separately: catalog registry, Lance manifests, async jobs, and object storage. Specify Query Dynamo actions/tables as read-only manifest access and explicitly exclude registry access.
3. Document migration as an operator-controlled copy/cutover; do not add a runtime fallback to the old shared table.
4. Run the Kubernetes selector, doc checks, fmt, and strict CLI clippy.

Task 4: Verify, review, and ship

Files:

• Create: verification/issue-124-query-manifest-authority.md

1. Run all completion selectors and mise run spec-lifecycle specs/issue-124-query-manifest-authority.spec.md.
2. Run affected strict clippy/check, git diff --check, then mise run gate.
3. Record exact evidence and request independent correctness/security and deployment review; resolve every P0/P1.
4. Commit refactor(cli): isolate query manifest authority (#124), push, open the PR, merge only after APPROVE, and verify #124 closes.