Evaluate the Claude Agent SDK as an alternative/augment to shell hooks

[wimaan3]

Context

Today borromeanRings integrates via .claude/settings.json shell hooks that shell out to the gate. The Claude Agent SDK (Python/TS) offers in-process hooks, a canUseTool permission callback, headless/CI-native operation, session resume, and multi-provider support (Bedrock/Vertex). See findings in #70.

Scope

• Prototype the gate as an SDK-driven harness (in-process PreToolUse/PostToolUse/Stop + canUseTool).
• Compare against the shell-hook approach on: robustness, statefulness, testability, latency/cost, CI-headless operation, portability.
• Decide: migrate, augment (SDK for CI + hooks for interactive), or stay.

Acceptance criteria

• A working SDK prototype running the existing gate.
• A comparison matrix (shell hooks vs SDK) against the criteria above.
• ADR with the recommendation; if adopting, a migration plan that keeps substrate-neutrality.

Relates to: #70, #4 (instrumentation), #54 (identity guard).