Project domestication: App-Server-first control-plane reboot

[sguzman]

Summary

Domesticates Brazen after the long pause and narrows the project around one product: a visible local control plane for ChatGPT Web <-> Codex App Server.

A major scope-changing fact was incorporated during the reboot: OpenAI now exposes the Codex harness through the official Codex App Server, which is the first-class rich-client integration boundary. Brazen therefore should not recreate Codex-native runtime machinery.

What changed

• adds AGENTS.md with the human -> director -> grunt operating model
• adds docs/PRODUCT.md and docs/CURRENT_STATE.md
• replaces the parallel-roadmap execution model with one active roadmap
• specifies Milestone 1 in docs/milestones/001-chatgpt-codex-control-plane.md
• records control-plane-first in ADR 0001
• records App Server adoption / scope reduction in ADR 0002
• stages the next Codex re-entry evidence pass in .ai/TASK.md
• adds .ai/REPORT.md and a revised .ai/BLOCKERS.md
• rewrites the root README around the reduced product boundary
• updates .gitignore to stop new local logs and patch/recovery artifacts from accumulating

App Server changes the project boundary

Codex App Server owns

• Codex agent loop
• thread lifecycle/persistence
• auth/config/model discovery
• tool execution/extensions
• turn/item event lifecycle
• streaming events
• diffs
• approval requests
• Codex-native protocol semantics

Brazen owns

• making ChatGPT Web routable
• explicit ChatGPT conversation <-> Codex thread bindings
• cross-surface routing
• allow / hold / block policy
• visible queues
• durable provenance/history
• project association
• topology/health visualization
• wire traffic/queue/failure metrics
• causal reconstruction after failures

The core architectural rule is now:

Make ChatGPT routable. Adapt App Server. Build the wire.

Milestone 1 target

• reliable ChatGPT page in Brazen
• multiple ChatGPT conversations
• thin Codex App Server client/adapter
• explicit persisted bindings
• live visual topology and health
• inspectable routed traffic and bulk history
• messages/events per minute, bytes per minute, queue/failure metrics
• independent allow / hold / block controls per direction
• durable causal history so the human no longer has to be the message bus

Important non-change

This PR does not delete existing recovery artifacts, old roadmap files, or broad platform code. Those are retained until the staged re-entry task classifies what is actually safe to remove or dormant.

This PR also makes no build/test claims. .ai/TASK.md intentionally makes build/test/launch/ChatGPT/App-Server compatibility verification the first implementation task after review.

Direct attachment to the already-running Codex Desktop process is explicitly not a Milestone 1 dependency. A Brazen-owned App Server process is acceptable.