refactor(host): reject point dispatch at the protocol boundary

[hqhq1025]

Why

Maka's production Computer Use action space is semantic-first and shared across platforms. The native executor protocol should therefore make coordinate mutation unavailable even to an older host that still knows the dispatch.point request shape.

Removing the method outright would turn a stable domain refusal into protocol-version ambiguity. Executing it would preserve a second mutation boundary that Windows and future platforms could accidentally implement differently.

What changes

• advertise an empty host.hello.capabilities.pointActions
• retain dispatch.point decoding as a compatibility endpoint
• return a stable unsupported_action refusal with path: none and no verification
• refuse before resolving or consuming a snapshot, inspecting the target, selecting a path, posting input, or performing post-action observation
• remove the unused point path selector and point-success/live tests
• preserve PID-bound native events only as internal implementations of already bound semantic actions such as element scrolling and WebContent activation
• document the distinction between the semantic maka.cu/2 boundary and this repository's separate legacy MCP/CLI coordinate APIs

Validation

• swift test --filter HostDispatchTests (24/24)
• swift test --filter HostProtocolTests (84 passed, 1 opt-in live test skipped)
• swift test (291 passed, 25 opt-in live tests skipped, 0 failures)
• git diff --check

Companion model/runtime action-space PR: apache/maka#4497

Related: apache/maka#4318

[Astro-Han]

LGTM!