CDP

• Collapsed the CDP snapshot surface to DOM-only. Removed cdp_take_ax_snapshot and the paired a<N> UID namespace. The native macOS take_ax_snapshot (still a<N>) and browser-side cdp_take_dom_snapshot / cdp_find_elements (always d<N>) now split cleanly — no more overlapping "which snapshot do I take?" for CDP. Breaking: existing callers that invoked cdp_take_ax_snapshot must switch to cdp_find_elements (for targeted lookups) or cdp_take_dom_snapshot (for the full page).
• Parallelized DOM.describeNode resolution. cdp_take_dom_snapshot and cdp_find_elements previously did three sequential CDP round trips per element (get ref, describe, release) — for 500 elements that was ~1500 serial round trips. The per-element chain now runs through futures::join_all, pipelining over the single CDP WebSocket.
• include_snapshot auto-appends capped at 100 nodes. cdp_click / cdp_hover / cdp_fill / cdp_press_key with include_snapshot=true previously appended a full 500-node DOM snapshot; they now append a 100-node snapshot. The user-facing cdp_take_dom_snapshot(max_nodes=500) default is unchanged.
• cdp_wait_for snapshot is now opt-in. Added an include_snapshot flag (default false). On success the response is now a one-line "Text appeared after Xms: [...]" header unless include_snapshot=true, in which case a 100-node DOM snapshot is appended after the header. Breaking: callers that relied on cdp_wait_for implicitly returning a snapshot must pass include_snapshot=true.
• cdp_element_at_point description corrected. Now accurately documents that the tool always returns backend_node_id and only carries a d-prefixed UID / role / name when the current DOM snapshot already contains the hit-tested node.