fix(computer-use): vision capture returns an image on cua-driver >=0.5.x

[OutThisLife]

Problem

computer_use capture(mode='vision') returned a 0x0 result with no image on every call:

{"mode":"vision","width":0,"height":0,"elements":[],"total_elements":0}

som and ax modes worked fine, which masked the regression.

Root cause

The vision branch called a screenshot MCP tool that cua-driver dropped in 0.5.x — full-window PNG capture was folded into get_window_state (its description even notes "Also captures a PNG screenshot of the specified window"). The driver replies Unknown tool: screenshot, so the flattened result carried no image part, images came back empty, png_b64 stayed None, and capture returned 0x0.

Verified against cua-driver 0.5.1:

$ cua-driver call screenshot '{"window_id":...}'
Unknown tool: screenshot
$ cua-driver list-tools | grep screenshot   # (no match)

som/ax were unaffected because they already use get_window_state.

Fix

Route vision by capability (tools/computer_use/cua_backend.py):

• driver advertises screenshot (older builds) → use it (cheapest, no AX walk)
• otherwise → call get_window_state but discard the AX tree/elements, returning only the PNG so vision stays free of element noise (that's the whole point of vision mode)
• capabilities not yet discovered → try screenshot, fall back to get_window_state on an empty image, so the path self-heals on any driver version

Also added _image_from_tool_result, which pulls the PNG from either an MCP image content-part or structuredContent.screenshot_png_b64. Applied on the som path too, so the image won't silently drop on driver builds that deliver it via structuredContent instead of a content part.

Verification

• Live (cua-driver 0.5.1): vision → 1568x954, 0 elements; som → image + 527 elements ✅
• Unit: all four routing cases (0.5.x no-screenshot, older with-screenshot, caps-undiscovered fallback, som unchanged) + the _image_from_tool_result shape handling
• tests/computer_use — 24 passed

[github-actions]

🔎 Lint report: bb/fix-vision-capture vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11461 on HEAD, 11461 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 6038 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[alt-glitch]

Duplicate of #39262 (the earliest open fix PR for #39242) — both route computer_use(capture, mode='vision') through get_window_state instead of the dropped screenshot MCP tool on cua-driver 0.5.x. Same cua-capture cluster as #44715. Marking duplicate so reviewers can compare against the canonical PR; the capability-fork + _image_from_tool_result structuredContent handling here is a useful refinement worth folding in.