[liubing0806]

Add an explicit kubernetes-ha mode so multiple control-plane Pods can concurrently serve ordinary Claude Docker Messages requests using shared PostgreSQL state and a pinned TLS node protocol. The existing kubernetes deployment remains single-instance. Slot payloads reuse the existing prebuilt kernel and CLI packages.

This is a scoped implementation of #648 and remains Draft. It depends on the PostgreSQL work in #649, which is still open; that dependency is included in this branch's diff.

Behavior

• PostgreSQL stores authentication/session state, routing/sticky state, FIFO turns, key/account/VM admissions, rate limits, durable jobs and idempotent request settlements. Each replica handles requests independently.
• Node allocations, initial credential publication and launch require private grants, approved policies and original receipts. Refresh tokens stay in the node's unmounted canonical directory; guests receive access-token-only projections.
• A lost controller does not release an unknown execution by timeout. Recovery requires loss of the original owner lock and the original node/command/execution identity's durable completion proof. Missing usage or pricing evidence retains the reservation.
• Recovery scanning, background ownership, maintenance and shutdown share the same persisted authority. HTTP/SSE drain accepted work; idle WebSockets close with 1001 during draining.
• deploy/k8s-ha provides two replicas, required node anti-affinity, Pod-local runtime volumes, restricted security settings, readiness/draining, PDB and an optional HPA. The offline bootstrap CLI verifies an already migrated business schema before private initialization or node actions.

Supported scope

Ordinary Claude Docker: Anthropic Messages and Claude Chat/Completions conversion over HTTP/SSE. GPT/Codex, API backends, Search, native/KVM, terminals and legacy management/import/restore paths are rejected in HA mode. Claude Responses retains its existing 400 protocol_not_allowed; successful WebSocket model inference is not supported by this implementation.

Validation

• Final code ff4e43486fa98af21aee366f8c3779202e81e1d0: fixed-archive CLI/database checks 25/25, 779-file format check and changed-file syntax checks passed. Final independent read-only Codex session found no new demonstrable P0/P1/P2 in the final delta. Review reports and earlier fixes are committed.
• 48aa2a28: 50/50 targeted checks and 9/9 real two-node Kubernetes smoke scenarios passed. Running rootfs layers and eight runtime source files were compared with the frozen source.
• f53e005f: 73/73 targeted checks; 2→3→2 scaling and rolling update with 45/45 HTTP/SSE requests returning 200; HPA 2→3 with a test-only CPU target of 1% and 31/31 requests returning 200; CRI exit 137 followed by recovery from the original node receipt and exactly one settlement. WebSocket upgrade/error/drain/bounded reconnect was checked, including a transient reconnect error.
• These cluster tests used actual PostgreSQL, Kubernetes and Go TLS node processes with a synthetic kernel/provider and fake Engine. A separate actual prebuilt ELF test verified AT projection reload without restarting the kernel/CLI. Real model inference, real provider RT refresh, existing production nodes and Ingress runtime behavior remain unverified.
• Full-suite checkpoint 4de403f9: unit 4339 passed / 24 failed / 10 skipped; E2E 117 passed / 17 failed / 1 skipped. Seven unit and all 17 E2E failures match the prior verified baseline. The additional 17 unit failures were missing test database initialization; the runner fix was separately rechecked 18/18, not represented as a full-suite rerun. Go race: 299 passed / 1 auxiliary skip; subsequent changes do not modify worker sources.

See the verification report for exact SHAs, image identities, review sessions, failure summaries and validation boundaries. No production deployment was performed.

Refs #648