Execution identity metadata, version 2.0

SV2-IDN-02 extends AgentRecord with execution_id, generation, runtime_backend and runtime_target. Canonical names continue to come from NameRegistry. A seat, its memories and its task lineage survive changes of execution identity or canonical process name. These records provide identity metadata only; they carry no grant and are not evidence of worker liveness or permission to mutate a task.

RedisStore.start_execution admits an incarnation before runtime creation. Its input is a named AgentRecord without an execution ID or generation. The caller supplies the exact previous_execution_id for a replacement, or an empty predecessor for the seat’s first tracked attempt. The store allocates a UUID and increments the seat’s execution generation in one watched transaction. Generations describe executions of a seat; task ownership generations and controller epochs belong to the later fencing package.

A Kubernetes target contains pod_namespace, pod_name and an optional pod_uid. The UID is bound once through put_agent after runtime creation or reconciliation; all other target fields are immutable. Local targets may carry server_id, process_namespace and a positive pid. A missing local target preserves compatibility for metadata admission before the local runtime has started. Different Pods may share a display name; execution lookup accepts only the immutable execution ID. Reusing any previously recorded Pod UID is refused. UUID collisions, changed incarnation fields and stale predecessor or update requests are also refused. execution_identity_conflicts_total counts refusals per swarm without granting authority.

RedisStore.execution retrieves an exact attempt. RedisStore.executions returns the seat’s ordered history. RedisStore.execution_occupants reconstructs the latest recorded incarnation per seat from persisted attempts, even when the current agents projection has been lost. This projection includes the last observation, including terminal states; the controller must consult runtime observations and execution authority before taking actions. Attempt records survive drop_agent, controller reconnection and the existing export/restore snapshot path. Latest attempt metadata can change through put_agent; superseded attempt records cannot be updated by that interface.

The preexisting local writer remains available for AgentRecord rows with default metadata. A remote record or a name already tracked in execution history must use admitted identity. Runtime adapter wiring and authenticated grants are separate dependent work packages; this package creates no Pods, processes or task leases.

Rollback: stop distributed launches before reverting the writer. Keep executions and seat memory as authoritative data. Old local readers can project the preceding AgentRecord fields; current readers also accept records that lack the added fields. The isolated rollback rehearsal checks this projection and verifies export/restore preserves all three attempts. Revert code through Git; do not delete attempt records to simulate rollback.

The agents hash remains a projection with the preceding record shape. New readers obtain additive identity from the executions hash in a consistent read transaction. This preserves the preceding strict dataclass decoder without requiring an old binary to understand the four added fields. Local writes watch both projection and execution history, so concurrent admission cannot be overwritten by an unadmitted legacy write.