feat: governed agent architecture (ADR-0036)

[nerdsane]

Summary

• Governed Agent Architecture (ADR-0036): Implements a full agent runtime as governed Temper specs — session tree storage (JSONL append-only with branching), context compaction, two-loop steering (CheckSteering/ContinueWithSteering/FinalizeResult), and event streaming
• New entity types: AgentSoul, AgentSkill, AgentMemory, ToolHook, CronJob, CronScheduler, HeartbeatMonitor, Channel, AgentRoute, ChannelSession — each with IOA specs, CSDL schemas, Cedar policies, and WASM integration modules
• Subagent + coding agent management: Entity-to-entity spawning via OData, CLI process management (Claude Code, Codex) via sandbox bash
• Shared libraries: session-tree-lib (JSONL tree data structure), wasm-helpers (TemperFS I/O, field extraction, URL resolution)
• Channel system (os-apps/temper-channels/): Webhook/Discord/Slack channel entities with route resolution, session tracking, and reply delivery
• Code review fixes: Extracted duplicate helpers into shared crate, server-side OData filtering for ChannelSessions, real timestamp comparison in heartbeat scan, Cedar policy for agent self-memory management

New OS Apps

App	Entities	WASM Modules
temper-agent	TemperAgent (extended), AgentSoul, AgentSkill, AgentMemory, ToolHook, CronJob, CronScheduler, HeartbeatMonitor	llm_caller, tool_runner, sandbox_provisioner, steering_checker, context_compactor, heartbeat_scan, cron_trigger, cron_scheduler_check, cron_scheduler_heartbeat, heartbeat_scheduler, coding_agent_runner
temper-channels	Channel, AgentRoute, ChannelSession	channel_connect, route_message, send_reply

Test plan

• cargo test --workspace passes (all gates green)
• cargo clippy clean
• rustfmt check passes
• Readability ratchet within baseline
• E2E proof run against live server (proof harness script at scripts/temper_agent_e2e_proof.py)

🤖 Generated with Claude Code