docs + server: ACP research-state schema and transport adapter skeleton

[n24q02m]

Summary

Dr. Claw research sessions accumulate structure that a linear transcript cannot express: open questions, evidence for and against claims, blockers, and append-only decision records. This PR adds:

1. docs/acp-research-state.md — a wire format for research state based on the Agent Client Protocol (ACP v1): four object types (task-graph nodes, typed edges, decision log entries, and claim→evidence expressed as typed edges) plus two JSON-RPC methods (session/record_research, session/get_research_state) over the standard ACP stdio transport.
2. server/acp-transport.js — a dependency-free skeleton for hosting any ACP-speaking agent as a provider (stdio subprocess, JSON-RPC client side, session/update event surface, and an in-memory research-state mirror that answers the agent's session/record_research requests).

The goal is interoperability without code merges: any agent that speaks ACP (e.g. an agent daemon exposing sessions over ACP) can run inside Dr. Claw, and any ACP client can render a Dr. Claw research session's task graph and decision log in the same format. The identical schema is being proposed upstream to ACP as a "Research Process Extensions" RFD.

Notes

• No new dependencies — Node core modules only (child_process, readline); plain JS ES modules per server/AGENTS.md.
• The adapter is a skeleton: it is not wired into server/index.js yet. I left the wiring point for discussion (provider picker + spawn-command config), since it touches UI and config plumbing.
• Original work authored from the public ACP specification; no third-party code or text is included.

Testing

• node --check server/acp-transport.js passes.
• A local two-peer round-trip (two linked createAcpPeer instances over in-memory streams) verifies: initialize handshake, unknown-method error (JSON-RPC -32601), a session/record_research batch (upsert node, add edge, record decision) acknowledged with a bumped stateVersion, and peer notification delivery.
• The module is intentionally not imported anywhere, so no existing behavior changes.

Questions for maintainers

1. Should the ACP provider be a first-class entry in the provider picker with a configurable spawn command, or an extension of the existing CLI-agent adapters?
2. Preferred persistence for the research-state mirror (per-session file vs. the existing session database)?
3. Interested in this direction at all — happy to iterate or split into smaller steps.

[n24q02m]

Withdrawing this PR — superseded by internal direction change. Thanks for the review effort.