feat(validator): R1 MVP P0 — OcrRoiChecker + orchestrator (flag OFF default)

Package core/validation/ minimal :
- result.py : Verdict, FailureCategory, ValidationResult
- pixel_diff_checker.py : wrapper de ReplayVerifier.verify_action
- ocr_roi_checker.py : ROI 80px autour du clic, détecte WRONG_APPLICATION
  via SUSPECT_TOKENS (edge/https/explorateur de fichiers/…)
- orchestrator.py : Validator dispatch action_type → checkers + agrégation

Wiring api_stream.py:3646 derrière RPA_VALIDATOR_V2_ENABLED (OFF par défaut).
Si verdict ≠ COMPLETE, override report.success=False et expose failure_category
dans result_entry. Zero régression flag OFF.

Tests :
- tests/unit/test_validator_v2.py : 13 tests (Checkers + Validator + sérialisation)
- tests/integration/test_validator_step10.py : 2 tests reproduisant le bug
  replay_sess_4c38dbb8 / act_raw_6c1432b3 (clic Enregistrer fait basculer
  vers Explorateur de fichiers) — Validator retourne WRONG_APPLICATION

Activation pour test live : RPA_VALIDATOR_V2_ENABLED=true

Cf. docs/recherche/SPEC_VALIDATOR_MATRICE.md, AXE_B2_DEEP_VALIDATOR.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

agent_v0/server_v1/api_stream.py

@@ -53,6 +53,42 @@ _replay_verifier = ReplayVerifier()
 _replay_learner = ReplayLearner()
 _audit_trail = AuditTrail()
 
+# Validator V2 (MVP P0) — flag RPA_VALIDATOR_V2_ENABLED, OFF par défaut.
+# Lazy init au premier appel : EasyOCR/docTR sont chargés à la demande.
+_validator_v2 = None
+
+
+def _get_validator_v2():
+    """Lazy init du Validator V2. Active uniquement si flag ON."""
+    global _validator_v2
+    if _validator_v2 is not None:
+        return _validator_v2
+    if os.environ.get("RPA_VALIDATOR_V2_ENABLED", "").lower() not in ("1", "true", "yes", "on"):
+        return None
+    try:
+        from core.validation import (
+            OcrRoiChecker,
+            PixelDiffChecker,
+            Validator,
+        )
+        pixel = PixelDiffChecker(_replay_verifier)
+        ocr_click = OcrRoiChecker(radius_px=80)
+        ocr_type = OcrRoiChecker(radius_px=120)
+        _validator_v2 = Validator(
+            checkers={
+                "click": [ocr_click, pixel],
+                "double_click": [ocr_click, pixel],
+                "right_click": [ocr_click, pixel],
+                "type": [ocr_type, pixel],
+            },
+            default_checkers=[pixel],
+        )
+        logger.info("[VALIDATOR_V2] activé (flag RPA_VALIDATOR_V2_ENABLED=ON)")
+    except Exception as exc:
+        logger.warning(f"[VALIDATOR_V2] init impossible: {exc}")
+        _validator_v2 = None
+    return _validator_v2
+
 # Nombre maximum de retries par action avant de déclarer un échec
 MAX_RETRIES_PER_ACTION = 3
 
@@ -3607,6 +3643,10 @@ async def report_action_result(report: ReplayResultReport):
     # Skip aussi la vérification serveur si l'agent a déjà géré la popup
     skip_verify = skip_verify or agent_handled_popup
     verification = None
+    # [VALIDATOR_V2] override conditionnel — flag RPA_VALIDATOR_V2_ENABLED.
+    # Si verdict ≠ COMPLETE, on force result.success=False et on expose failure_category.
+    validator_v2_result = None
+    validator_v2_failure_category = None
     if report.success and screenshot_after and not skip_verify:
         # Utiliser le screenshot_before envoyé par l'agent (Critic fiable)
         # Fallback sur le dernier screenshot stocké côté serveur
@@ -3617,7 +3657,35 @@ async def report_action_result(report: ReplayResultReport):
                 result_dict = {
                     "success": report.success,
                     "error": report.error,
+                    "actual_position": report.actual_position,
                 }
+                # === Validator V2 (P0 MVP) — derrière flag, sinon no-op ===
+                v2 = _get_validator_v2()
+                if v2 is not None:
+                    try:
+                        validator_v2_result = v2.validate(
+                            action=action_dict,
+                            result=result_dict,
+                            screenshot_before=screenshot_before,
+                            screenshot_after=screenshot_after,
+                            context={},
+                        )
+                        from core.validation import Verdict as _V2Verdict
+                        if validator_v2_result.verdict != _V2Verdict.COMPLETE:
+                            validator_v2_failure_category = (
+                                validator_v2_result.failure_category.value
+                                if validator_v2_result.failure_category else None
+                            )
+                            report.success = False
+                            logger.info(
+                                f"[VALIDATOR_V2] override success→False action={action_id} "
+                                f"verdict={validator_v2_result.verdict.value} "
+                                f"conf={validator_v2_result.confidence:.2f} "
+                                f"failure_category={validator_v2_failure_category} "
+                                f"reason={validator_v2_result.reasoning[:120]}"
+                            )
+                    except Exception as exc:
+                        logger.warning(f"[VALIDATOR_V2] échec: {exc}")
                 # Utiliser le Critic sémantique si l'action a un expected_result
                 expected_result = (original_action or {}).get("expected_result", "")
                 action_intention = (original_action or {}).get("intention", "")
@@ -3686,6 +3754,8 @@ async def report_action_result(report: ReplayResultReport):
             "actual_position": report.actual_position,
             "retry_count": retry_count,
             "verification": verification.to_dict() if verification else None,
+            "validator_v2": validator_v2_result.to_dict() if validator_v2_result else None,
+            "failure_category": validator_v2_failure_category,
             "resolution_method": report.resolution_method,
             "resolution_score": report.resolution_score,
             "resolution_elapsed_ms": report.resolution_elapsed_ms,