fix(cua): SoM click must focus the input — the core of the type-stall

[mercurialsolo]

The bottom layer of the type-stall

Live-verifying #951 showed typing still failed. Root cause is one layer
deeper: the SoM click path (cdp_click_at_point, which the runner uses for
unstructured clicks) dispatches el.click() via Runtime.evaluate — and a
programmatic el.click() fires a click event but does NOT focus a text
input (real mouse clicks focus on mousedown).

So:

1. Brain SoM-clicks the search box → el.click() → click fires, input not focused (activeElement stays body).
2. SoM swaps the action to a no-op WAIT, so fix(cua): tell the brain when a click focused an input — break the type-stall #951's env CLICK-branch probe never runs.
3. No focus → no focused_input → brain never told to type → stalls on wait()/Return.

This is the actual bottom of the audit's "type never lands."

Fix

• cdp_click_at_point: el.focus() focusable targets
  (INPUT/TEXTAREA/SELECT/contenteditable) before el.click(),
  mirroring native focus→click ordering; report focused on the result and
  _last_som_diag.
• runner: after a SoM click that focused a field, surface focused_input so
  the fix(cua): tell the brain when a click focused an input — break the type-stall #951 "input field focused — type your text now" feedback fires on the
  SoM path too — the brain types instead of waiting.

Tests

test_cua_input_focus_feedback.py — SoM payload focuses + reports focused.
215 adjacent tests green; ruff clean.

Verification

After merge + deploy I'll re-run the live typing test — expecting the brain to
finally type() into the focused search box. This is layer 3 of the type-stall
(brain-dead #948 → focus-feedback #951 → SoM-focus here); I'll confirm
empirically rather than assume.

🤖 Generated with Claude Code