[Verlintas]

What this adds

Phase 2 of the web visual channel:

• Full-page screenshots — web_research screenshot(full_page=true) uses WebView.capturePicture() to grab the whole document and crops it into up to six readable viewport segments. Longer pages are explicitly marked truncated. Scrolling + View.draw() was abandoned because the virtual-display WebView reports the new scroll offset but returns a blank draw after scrolling.
• Sensitive-page guard — screenshots are refused with screenshot_blocked when the page has a visible password field, a card/payment form, a Stripe/PayPal-like payment iframe, a checkout/payment URL plus payment text, or a CAPTCHA challenge. The model is pointed to snapshot / read / show_to_user instead.
• User-visible receipts — AgentEvent.ToolCallFinished carries ephemeral ToolImages. When the existing keepExecutionScreenshots setting is on, ChatViewModel imports them as assistant-message attachments, so the user can see exactly what the model saw. Assistant attachments are stripped from future model input, as before.
• Main-screen tools — take_screenshot and screen_ocr now return OCR plus the image when the model is a known vision family (or vision=true); vision=false keeps them text-only.
• Startup WebView mode — enableSlowWholeDocumentDraw() is enabled in Application.onCreate, before any WebView loads, so capturePicture() includes content below the first viewport.
• Debug tooling — generic DebugToolActivity for any registered tool, and DebugOcrActivity can OCR a file path.

Verification

API 35 emulator, devtest package; daily app untouched.

• 5-viewport local page: 5 images, 799×1600 each; OCR read ALPHA-111, BRAVO-222, CHARLIE-333, DELTA-444, ECH0-555.
• 9-viewport page: 6 images with a clear truncation note.
• full_page=true&vision=false: 5 OCR segments, images=0.
• password.html: screenshot_blocked: ... password field.
• payment.html (autocomplete=cc-number/cc-csc): screenshot_blocked: ... card/payment form.
• Single viewport chart page: still 720×1600 / 135 KB with OCR.
• take_screenshot with vision=true: 92 KB PNG + OCR + images=1.
• screen_ocr with vision=true: OCR + images=1 (entire-screen MediaProjection grant).
• Unit tests: WebScreenshotPolicy (allowed article, password/card, payment frame/page, CAPTCHA) and AgentEvent image propagation.
• test, lintDebug, assembleDebug, assembleRelease green.

Known limits

• Full-page is capped at six segments; very long pages need follow-up scrolling or read.
• capturePicture() may miss video/WebGL/SurfaceView layers on some WebView builds.
• enableSlowWholeDocumentDraw() raises memory/time for very long pages.
• Main-screen screenshots still rely on the user's MediaProjection grant and may contain notifications; the tool description warns against password/payment screens.