Part of #3896 phase 1.

Summary

Add the pure planning engine that converts validated channel manifests into a serializable SandboxMessagingPlan. This is architecture-only work: no production workflow should use the compiler yet.

Scope

• Finalize SandboxMessagingPlan so it contains top-level channels, credentialBindings, networkPolicy, agentRender, buildSteps, stateUpdates, and healthChecks.
• Add compiler context types for sandbox name, agent, workflow, selected channels, configured channels, disabled channels, persisted state, config inputs, and credential availability.
• Add pure engines: input-resolver.ts, credential-binding-engine.ts, policy-resolver.ts, agent-render-engine.ts, state-update-engine.ts, build-step-engine.ts, and health-check-engine.ts.
• Add ManifestCompiler.
• Add tests for plan JSON round-trip, no secrets, no function-valued fields, and deterministic output.

Acceptance Criteria

• Telegram, Discord, Slack, WeChat, and WhatsApp manifests compile into stable SandboxMessagingPlan fixtures.
• Provider names and placeholders match current behavior.
• Policy plans match current channel policy intent, including Hermes WeChat alias behavior.
• OpenClaw and Hermes render fragments are represented in the plan.
• Build args, build files, and rebuild hydration are represented without applying them.
• Health checks are planned before any lifecycle success can be reported.
• ManifestCompiler contains no channel-specific branches like if channel === "wechat".
• Compiler and engines do not import production side-effect modules.

Non-goals

• Do not write registry or session state.
• Do not create providers.
• Do not patch Dockerfiles.
• Do not modify OpenClaw or Hermes config files.