native: experimental Agent-SDK headless backend (Track B)

[itomek]

Part of #125. Experimental (Track B) — optional, behind a config flag.

Problem

The JSONL+hook layer is a bridge over PTY sessions. The Agent SDK offers natively structured sessions (canUseTool for permissions, typed message stream). The CodeBackend interface (backend/src/backends/types.ts:85-166) is backend-agnostic — validate that an SDK-based backend can slot in.

Outcome

A HeadlessBackend implementing CodeBackend via @anthropic-ai/claude-agent-sdk: plain-text task:output, canUseTool → task:prompt, resume support — created per-task behind a config flag.

Acceptance criteria

• Create/drive/stop/resume a task on the SDK backend via test-cli
• Transcript and permissions arrive structured natively (no JSONL tail, no hooks)
• resizeTask no-ops cleanly; state lifecycle maps correctly
• Documented feature-gap comparison vs the PTY backend (slash commands, plan mode, compaction visibility) as a comment on this issue
• Existing PTY backends untouched; flag off by default

Scope & expectations

• Where: new backend/src/backends/headless-backend.ts, config-store flag
• Constraints: SDK is pre-1.0 — pin the version; no changes to CodeBackend interface without discussion

How to verify

CLI — test-cli session transcript on the SDK backend including a structured permission answer.