UI TestEngine: pressButton/writeValue return status string (empty=OK)

[Grantim]

Summary

PR #5961 conflated "widget is disabled" with "caller error" by turning the disabled case into an Expected failure. That broke the internal test_all_scens scenario runner — the Python binding's expectedValueOrThrow raised inside the GUI-thread callback and the process landed on std::terminate (exit 6). #5968 narrowed the blocked heuristic but left the silent-to-error semantic intact.

This PR cleanly separates the two axes:

• Control::pressButton / Control::writeValue<T> now return Expected<std::string>. Empty string on success (click / write simulated). Non-empty string ("disabled[: <reason>]", same vocabulary as composeStatus) when the widget was drawn disabled — in that case the simulation is a silent no-op. unexpected(...) is reserved for hard errors only: path not found, wrong entry kind, value out of range / not in allowedValues.
• The Python binding (mrviewerpy.uiPressButton / uiWriteValue) logs non-empty status via spdlog::warn and does not throw on disabled — pre-UI TestEngine: expose disabled/blocked state on entries #5961 silent contract restored.
• The MCP handler translates non-empty status into a thrown std::runtime_error with the same "pressButton <path>: disabled: <reason>" wire-format as before, so LLM-facing behavior is unchanged.

readValue, listEntries, listAllEntries, and the TypedEntry::status field are untouched.

Test plan

• Debug|x64 build clean locally.
• MCP: ui.pressButton on a disabled entry (QuickAccess/Undo) → thrown error "pressButton QuickAccess/Undo: disabled: Nothing to undo." (wire format byte-identical to pre-fix).
• MCP: ui.pressButton on an enabled entry → empty-JSON success.
• MCP: ui.writeValueReal on an enabled float field → empty-JSON success.
• MCP: ui.pressButton on a non-existent path → hard-error propagates through unexpected(...) as before.
• Python path + the previously-failing test_all_scens scenario — left for CI.

🤖 Generated with Claude Code