fix: wire permission_response WS frames to the interactive-agent-layer

[jlunder00]

Summary

• The keystone fix for the permission approve/deny loop (brief 1d). bot_chat's mid-session race only ever inspected the first non-stop client frame per turn (logged a warning, then fell through to just awaiting the session) — silently dropping permission_response frames entirely. This is why the approve/deny loop never reached the layer end-to-end.
• LayerClient.respond_to_permission(request_id, approve) — POSTs to /permission/{request_id}/respond, matching the real PermissionRespondRequest schema ({"approve": bool}) in interactive_agent_layer/server.py.
• api/routes/bot.py::_respond_to_permission maps the frontend's decision ("approve"/"deny") to the layer's boolean approve field; failures are logged and swallowed (the gate's own permission timeout is the backstop).
• Restructured the mid-session race into a loop that re-arms recv_task after each frame, so multiple mid-session frames (e.g. two permission_responses, or a stop after one) are all consumed within a single turn — not just the first.

Test plan

• New: tests/interactive_agent_layer/test_client.py — 3 tests for LayerClient.respond_to_permission (approve, deny, 404-raises) against a real ASGI-backed layer app.
• New: tests/api/test_bot_ws_permission_response.py — 6 integration tests: single approve round trip, deny→approve=False mapping, TWO permission_responses consumed in one turn, layer POST failure logged+swallowed (session still completes), missing request_id dropped without crashing, unrelated frame types still dropped+warned (regression guard).
• Full existing suite (test_bot_ws.py, test_ws_event_types.py, test_agent_dispatch_ws.py, tests/interactive_agent_layer/) green — stop-cancels-session and disconnect-cancels-session behavior unaffected.
• Full repo suite: 992 passed, 740 skipped (DB-dependent), 0 failed.

🤖 Generated with Claude Code