Per-run state threaded through Hourglass.Workflow.Evaluator.evaluate/3.

Across activations:

• next_global_seq — monotonic SDK seq counter; persists across activations (Temporal requires monotonic seqs per workflow lifetime).
• pending_resolvers — %{global_seq => command_id} registered when a command was scheduled. The next activation's resolve_activity{seq} uses this to map back to the command_id.
• command_id_to_seq — inverse of pending_resolvers. Lets re-execution look up "have I seen this command_id before? what seq did it get?"
• resolved_results — %{global_seq => decoded_value} populated from prior resolve_activity jobs delivered on this or earlier activations.
• input — the decoded initialize_workflow first argument; set once.
• workflow_module — the workflow module bound to this run (e.g. MyApp.Workflows.IngestSource). Set by Hourglass.Workflow.Evaluator.evaluate/3 on the first activation and persisted across cache round-trips so subsequent activations (e.g. resolve_activity deliveries that don't carry initialize_workflow) can be routed to the same module by the worker's workflow type resolver.

Per-activation (rebuilt each call):

• commands — newest-first list of {command_id, command_term} issued during this activation's re-execution. Sorted + assigned seqs at the end of evaluate/3.

• result — nil | {:completed, term} | {:failed, term} | :evict determines the completion shape.

• child_count — counter that replaces the expected_children MapSet. The pure-function evaluator runs async/1 inline, so it has no child pids to track; the counter only exists for diagnostics.