feat(runtime): local workflow orchestration (closes #23)

[leo-aa88]

Summary

Implements issue #23: MVP local runtime that loads YAML from a project root, applies Environment overrides, validates workflow input before persisting a run, then drives engine.Executor.

API

• runtime.WorkflowRunner / runtime.WorkflowRunOptions — thin interface in internal/runtime so cmd stays decoupled from wiring details.
• local.Runtime + local.NewRuntime — ProjectRoot + state.RuntimeStore.
• local.Runtime.ExecuteWorkflow — project.LoadProject → ApplyEnvironment (agents + policies per design doc section 7.6) → parse InputJSON → engine.ValidateWorkflowInput → util.NewRunID when RunID empty → StartRun → run.started trace → Executor.Run → run.finished trace.

Environment overrides (MVP)

• Agents: optional model; constraints merged with non-zero / explicit flags (timeout, iterations, temperature, structured output).
• Policies: execution limits merged (wall clock, cost ceiling, structured output flag).

Unknown agent/policy names in an environment override return an error.

Engine change

• Exported engine.ValidateWorkflowInput so the runtime can reject invalid input before StartRun (malformed JSON is rejected earlier in the local runner).

Tests

• SQLite: full demo workflow (tool + mock agent) with EnvironmentName: staging; asserts succeeded run, trace includes run.started / run.finished.
• Invalid JSON: no runs row.
• Invalid schema (valid JSON, wrong shape): no runs row.
• ApplyEnvironment: reviewer timeoutSeconds = 99.
• Auto run ID when RunID omitted.

Verification

• make fmt
• make vet
• make test (-race)

Follow-ups

• Wire agentctl run --project to local.Runtime.
• Richer policy merge / override semantics if specs evolve.

Closes #23

Made with Cursor