fix(computer-use): route cua screenshots through aux vision

[im-khang]

Summary

• make the Cua computer-use backend enumerate all windows, then prefer on-screen candidates, so capture/focus still work when cua-driver marks a capturable app window off-screen or on another Space
• route som, vision, and ax captures through get_window_state with explicit capture_mode, and read dimensions/tree data from structuredContent
• create the aux-vision cache directory before writing captured screenshots, and ignore failed vision_analyze envelopes instead of treating their error text as valid analysis

Bug fixed

computer_use with cua-driver could return 0x0 geometry, use the wrong screenshot path for vision, miss app-filtered windows, or fall back to multimodal screenshot output when aux vision routing failed because $HERMES_HOME/cache/vision did not exist. For text-only main models, that fallback caused the active provider to reject the tool result image.

Verification

• python -m pytest tests/tools/test_computer_use_capture_routing.py tests/tools/test_computer_use.py -q -o 'addopts='
• local live smoke on macOS with Calculator:
  • capture app="Calculator" mode="vision" returned width=460 height=816, vision_analysis_routed_via=auxiliary.vision, no error
  • capture app="Calculator" mode="som" returned width=460 height=816, vision_analysis_routed_via=auxiliary.vision, no error

[Morad37]

Clean unification of the three capture paths. One thing I noticed: the _split_tree_text fallback path at line ~442 still runs even when structuredContent provides tree_markdown, so you're parsing the AX tree twice on every som/ax capture that has structured content. Moving the parsed_tree parse inside the if not structured_tree branch would save the regex work on cua-driver versions that already ship the structured tree.

[teknium1]

Thanks for the focused capture-routing work. The auxiliary-vision and cache-directory portions are now already present on current main: tools/computer_use/tool.py:778-832 creates the cache directory and returns the auxiliary.vision text result. Current main also has capability-aware vision capture routing at tools/computer_use/cua_backend.py:1373-1418, with regression coverage at tests/tools/test_computer_use.py:2360-2407.

Problems

• The remaining change is not a clean replay: this PR changes window discovery to include off-screen windows, while current main explicitly defines both capture() and focus_app() as on-screen selection paths (tools/computer_use/cua_backend.py:1247-1264, :1738-1776).
• Replacing current routing with unconditional get_window_state would bypass the current legacy-screenshot capability path (tools/computer_use/cua_backend.py:1387-1418).

Suggested changes

• Rebase the remaining off-screen-window case onto current CuaDriverBackend and prove it with a targeted driver reproduction plus mixed-inventory tests, preserving the current capability and CLI-fallback paths.

Automated hermes-sweeper review.