feat: Self-Healing Observation Loop + Visual Verification 2.0

[Delqhi]

Problem

When a click action fails silently (DOM unchanged, page looks the same), the script blindly continues and eventually crashes or produces wrong results. There is no feedback mechanism to detect "nothing happened".

Evidence

• Xiumu: Chrome Auto Browse Reality Check — "You have to watch it... its more like babysitting an easily distracted robot"
• Medium: 6 Failure Envelopes — "They fail inside a believable range"

Solution (Implemented)

Self-Healing Loop (click_ref_checked)

1. Take screenshot + DOM snapshot BEFORE click
2. Execute click via CDP (with Human Entropy)
3. Take screenshot + DOM snapshot AFTER click
4. Compare DOM diff (addedCount, removedCount)
5. Compare screenshots (Visual Verification 2.0 — Base64 length diff proxy)
6. If BOTH are zero → trigger JS DispatchEvent fallback automatically
7. Re-verify after fallback

Visual Verification 2.0

• Compares Base64 encoded PNG data length before/after
• A real UI change modifies compressed image size by >0.1%
• If diff_ratio < 0.001 → "KEINE visuelle Änderung" → trigger self-healing

Files Changed

• real_bridge_prolific.py — click_ref_checked(), visual_verification_2_0(), page_diff()

Acceptance Criteria

• Silent click failures are detected within 2 seconds
• JS fallback fires automatically without human intervention
• Visual verification correctly distinguishes real UI changes from no-ops
• Screenshots saved to /tmp/opensin-bridge-prolific/ for every action

[Delqhi]

Implementation note from cloud execution lane:

• Reported outcome: self-healing observation logic was moved into a dedicated runtime model, with before/after observation, DOM+visual scoring, fallback strategies, and proof retrieval via a dedicated surface.
• Reported validation: targeted bridge/runtime tests passed in the execution lane.
• Important repo-state note: branch/worktree hygiene remains necessary because the default local checkout contains unrelated modifications.