diff --git a/core/workflow/execution_compiler.py b/core/workflow/execution_compiler.py
new file mode 100644
index 000000000..ada04a5c6
--- /dev/null
+++ b/core/workflow/execution_compiler.py
@@ -0,0 +1,280 @@
+# core/workflow/execution_compiler.py
+"""
+ExecutionCompiler — Compile un WorkflowIR en ExecutionPlan.
+
+Pièce maîtresse de l'architecture V4.
+"Le LLM prépare et compile. Le runtime exécute."
+
+Le compilateur :
+1. Prend chaque étape du WorkflowIR
+2. Compile une stratégie de résolution pour chaque action (OCR > template > VLM)
+3. Définit les timeouts, retries, fallbacks et recovery
+4. Produit un ExecutionPlan déterministe et borné
+
+L'objectif : zéro VLM au runtime pour les cas normaux.
+Le VLM est un exception handler, pas le chemin principal.
+
+Le compilateur utilise :
+- Les données de l'enregistrement (crops, textes OCR) pour pré-compiler
+- L'historique d'apprentissage (ReplayLearner) pour choisir la meilleure stratégie
+- Le contexte métier (DomainContext) pour adapter les paramètres
+"""
+
+import logging
+import os
+import time
+import uuid
+from typing import Any, Dict, List, Optional
+
+from .workflow_ir import WorkflowIR, Step, Action
+from .execution_plan import (
+    ExecutionPlan, ExecutionNode, ResolutionStrategy, SuccessCondition,
+)
+
+logger = logging.getLogger(__name__)
+
+# Temps estimé par type d'action (ms)
+_ACTION_TIME_ESTIMATES = {
+    "click": 200,       # OCR lookup + clic
+    "type": 500,        # Frappe char-by-char
+    "key_combo": 100,
+    "wait": 0,          # Le duration_ms est dans l'action
+    "scroll": 200,
+}
+
+
+class ExecutionCompiler:
+    """Compile un WorkflowIR en ExecutionPlan.
+
+    Usage :
+        compiler = ExecutionCompiler()
+        plan = compiler.compile(workflow_ir, target_machine="VM_Win11")
+        plan.save("data/plans/")
+    """
+
+    def __init__(self, learning_dir: str = ""):
+        self._learning_dir = learning_dir or "data/learning/replay_results"
+
+    def compile(
+        self,
+        ir: WorkflowIR,
+        target_machine: str = "",
+        target_resolution: str = "1280x800",
+        params: Optional[Dict[str, str]] = None,
+    ) -> ExecutionPlan:
+        """Compiler un WorkflowIR en ExecutionPlan.
+
+        Args:
+            ir: Le WorkflowIR à compiler
+            target_machine: Machine cible (pour adapter les stratégies)
+            target_resolution: Résolution de la machine cible
+            params: Variables à substituer
+        """
+        t_start = time.time()
+
+        plan = ExecutionPlan(
+            plan_id=f"plan_{uuid.uuid4().hex[:8]}",
+            workflow_id=ir.workflow_id,
+            version=ir.version,
+            created_at=time.time(),
+            domain=ir.domain,
+            target_machine=target_machine,
+            target_resolution=target_resolution,
+            variables=params or {v.name: v.default for v in ir.variables},
+        )
+
+        # Consulter l'historique d'apprentissage
+        learned_strategies = self._load_learned_strategies()
+
+        # Compiler chaque étape
+        for step in ir.steps:
+            nodes = self._compile_step(step, ir, learned_strategies)
+            plan.nodes.extend(nodes)
+
+        # Statistiques de compilation
+        plan.total_nodes = len(plan.nodes)
+        plan.nodes_with_ocr = sum(
+            1 for n in plan.nodes
+            if n.strategy_primary and n.strategy_primary.method == "ocr"
+        )
+        plan.nodes_with_template = sum(
+            1 for n in plan.nodes
+            if n.strategy_primary and n.strategy_primary.method == "template"
+        )
+        plan.nodes_with_vlm = sum(
+            1 for n in plan.nodes
+            if n.strategy_primary and n.strategy_primary.method == "vlm"
+        )
+        plan.estimated_duration_s = sum(
+            _ACTION_TIME_ESTIMATES.get(n.action_type, 200) + n.duration_ms
+            for n in plan.nodes
+        ) / 1000.0
+
+        elapsed = time.time() - t_start
+        logger.info(
+            f"Compilation: {plan.total_nodes} nœuds en {elapsed:.1f}s — "
+            f"OCR={plan.nodes_with_ocr}, template={plan.nodes_with_template}, "
+            f"VLM={plan.nodes_with_vlm} (exception handler)"
+        )
+
+        return plan
+
+    def _compile_step(
+        self,
+        step: Step,
+        ir: WorkflowIR,
+        learned: Dict[str, str],
+    ) -> List[ExecutionNode]:
+        """Compiler une étape en nœuds d'exécution."""
+        nodes = []
+
+        for i, action in enumerate(step.actions):
+            node = self._compile_action(
+                action=action,
+                step=step,
+                action_index=i,
+                ir=ir,
+                learned=learned,
+            )
+            nodes.append(node)
+
+        return nodes
+
+    def _compile_action(
+        self,
+        action: Action,
+        step: Step,
+        action_index: int,
+        ir: WorkflowIR,
+        learned: Dict[str, str],
+    ) -> ExecutionNode:
+        """Compiler une action en nœud d'exécution avec stratégie de résolution."""
+
+        node = ExecutionNode(
+            node_id=f"n_{step.step_id}_{action_index}",
+            action_type=action.type,
+            intent=step.intent,
+            step_id=step.step_id,
+            is_optional=step.is_optional,
+        )
+
+        if action.type == "click":
+            # Compiler les stratégies de résolution pour ce clic
+            node.strategy_primary, node.strategy_fallbacks = self._compile_click_resolution(
+                action, step, learned,
+            )
+            node.timeout_ms = 10000
+            node.max_retries = 2
+            node.recovery_action = "escape"
+
+            # Condition de succès basée sur la postcondition
+            if step.postcondition:
+                node.success_condition = SuccessCondition(
+                    method="screen_changed",
+                    description=step.postcondition,
+                )
+
+        elif action.type == "type":
+            node.text = action.text
+            node.variable_name = action.text.strip("{}") if action.variable else ""
+            node.timeout_ms = 5000
+            node.max_retries = 0  # Pas de retry sur la frappe
+            node.recovery_action = "undo"
+
+        elif action.type == "key_combo":
+            node.keys = action.keys
+            node.timeout_ms = 3000
+            node.max_retries = 0
+            node.recovery_action = "undo"
+
+        elif action.type == "wait":
+            node.duration_ms = action.duration_ms or 1000
+            node.timeout_ms = action.duration_ms + 2000
+            node.max_retries = 0
+            node.recovery_action = "none"
+
+        elif action.type == "scroll":
+            node.timeout_ms = 3000
+            node.max_retries = 0
+            node.recovery_action = "none"
+
+        return node
+
+    def _compile_click_resolution(
+        self,
+        action: Action,
+        step: Step,
+        learned: Dict[str, str],
+    ) -> tuple:
+        """Compiler les stratégies de résolution pour un clic.
+
+        Ordre de priorité :
+        1. OCR exact (si texte connu) — 100ms, pixel-perfect
+        2. Template matching (si crop disponible) — 10ms, même interface
+        3. Position relative (si hint disponible) — instantané, fragile
+        4. VLM (dernier recours) — 2-5s, exception handler
+
+        Le learning peut réordonner si une stratégie a mieux marché avant.
+        """
+        primary = None
+        fallbacks = []
+
+        target_text = action.anchor_hint or action.target
+        learned_method = learned.get(target_text, "")
+
+        # Stratégie OCR — le texte visible est la meilleure ancre
+        if target_text:
+            ocr_strategy = ResolutionStrategy(
+                method="ocr",
+                target_text=target_text,
+                threshold=0.8,
+            )
+            # Si le learning dit que l'OCR marche pour cette cible, c'est la primaire
+            if not learned_method or learned_method in ("ocr", "som_text_match", "hybrid_text_direct"):
+                primary = ocr_strategy
+            else:
+                fallbacks.append(ocr_strategy)
+
+        # Stratégie template — le crop visuel de l'enregistrement
+        if action.anchor_hint:
+            template_strategy = ResolutionStrategy(
+                method="template",
+                target_text=action.anchor_hint,
+                threshold=0.85,
+            )
+            if learned_method in ("anchor_template", "template_matching"):
+                primary = template_strategy
+            else:
+                fallbacks.append(template_strategy)
+
+        # Stratégie VLM — exception handler (dernier recours)
+        vlm_description = action.target or step.intent
+        vlm_strategy = ResolutionStrategy(
+            method="vlm",
+            vlm_description=vlm_description,
+            threshold=0.6,
+        )
+        fallbacks.append(vlm_strategy)
+
+        # Si aucune primaire trouvée, utiliser le VLM
+        if primary is None:
+            if fallbacks:
+                primary = fallbacks.pop(0)
+            else:
+                primary = vlm_strategy
+
+        return primary, fallbacks
+
+    def _load_learned_strategies(self) -> Dict[str, str]:
+        """Charger les stratégies apprises (ReplayLearner)."""
+        try:
+            from agent_v0.server_v1.replay_learner import ReplayLearner
+            learner = ReplayLearner(learning_dir=self._learning_dir)
+            # Construire un mapping target → best_method depuis l'historique
+            strategies = {}
+            for outcome in learner._recent:
+                if outcome.success and outcome.resolution_method and outcome.target_description:
+                    strategies[outcome.target_description] = outcome.resolution_method
+            return strategies
+        except Exception:
+            return {}
diff --git a/core/workflow/execution_plan.py b/core/workflow/execution_plan.py
new file mode 100644
index 000000000..038a8ff46
--- /dev/null
+++ b/core/workflow/execution_plan.py
@@ -0,0 +1,251 @@
+# core/workflow/execution_plan.py
+"""
+ExecutionPlan — Plan d'exécution strict, borné et versionné.
+
+C'est ce que le runtime exécute. Pas d'improvisation — tout est pré-compilé :
+- chaque nœud a une stratégie de résolution primaire + fallbacks
+- chaque nœud a un timeout, un retry policy, une condition de succès
+- le VLM n'intervient qu'en exception handler (pas en chemin principal)
+
+Le runtime ne fait que : exécuter → observer → vérifier → suite ou fallback.
+
+Cycle : WorkflowIR → ExecutionCompiler → ExecutionPlan → Runtime
+"""
+
+import json
+import logging
+import time
+import uuid
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ResolutionStrategy:
+    """Stratégie de résolution visuelle pour un élément UI.
+
+    Pré-compilée — le runtime n'a pas besoin du VLM pour résoudre.
+    """
+    method: str                         # "ocr", "template", "position", "vlm"
+    target_text: str = ""               # Texte à chercher (pour OCR)
+    anchor_b64: str = ""                # Crop de référence (pour template matching)
+    zone: Dict[str, float] = field(default_factory=dict)  # Zone de recherche {x_min, y_min, x_max, y_max}
+    position_hint: str = ""             # "en haut à droite", "dans la barre des tâches"
+    vlm_description: str = ""           # Description VLM (dernier recours)
+    threshold: float = 0.8              # Seuil de confiance
+
+    def to_dict(self) -> Dict[str, Any]:
+        d = {"method": self.method}
+        if self.target_text:
+            d["target_text"] = self.target_text
+        if self.anchor_b64:
+            d["anchor_b64"] = self.anchor_b64[:50] + "..."  # Tronqué pour la lisibilité
+        if self.zone:
+            d["zone"] = self.zone
+        if self.position_hint:
+            d["position_hint"] = self.position_hint
+        if self.vlm_description:
+            d["vlm_description"] = self.vlm_description
+        d["threshold"] = self.threshold
+        return d
+
+    @classmethod
+    def from_dict(cls, d: Dict) -> "ResolutionStrategy":
+        return cls(**{k: v for k, v in d.items() if k in cls.__dataclass_fields__})
+
+
+@dataclass
+class SuccessCondition:
+    """Condition de succès d'un nœud — comment vérifier que l'action a marché."""
+    method: str = "screen_changed"      # "screen_changed", "title_match", "text_visible", "none"
+    expected_title: str = ""            # Titre fenêtre attendu après l'action
+    expected_text: str = ""             # Texte qui doit apparaître
+    description: str = ""               # Description pour le Critic VLM (exception handler)
+
+    def to_dict(self) -> Dict[str, Any]:
+        d = {"method": self.method}
+        if self.expected_title:
+            d["expected_title"] = self.expected_title
+        if self.expected_text:
+            d["expected_text"] = self.expected_text
+        if self.description:
+            d["description"] = self.description
+        return d
+
+    @classmethod
+    def from_dict(cls, d: Dict) -> "SuccessCondition":
+        return cls(**{k: v for k, v in d.items() if k in cls.__dataclass_fields__})
+
+
+@dataclass
+class ExecutionNode:
+    """Nœud d'exécution — une action à exécuter avec sa stratégie complète."""
+    node_id: str
+    action_type: str                    # click, type, key_combo, wait, scroll
+    intent: str = ""                    # Intention métier (pour le logging/audit)
+
+    # Résolution visuelle pré-compilée
+    strategy_primary: Optional[ResolutionStrategy] = None
+    strategy_fallbacks: List[ResolutionStrategy] = field(default_factory=list)
+
+    # Données de l'action
+    text: str = ""                      # Texte à taper
+    keys: List[str] = field(default_factory=list)
+    duration_ms: int = 0
+    variable_name: str = ""             # Si le texte est une variable
+
+    # Bornes d'exécution
+    timeout_ms: int = 10000             # Timeout pour cette action
+    max_retries: int = 1                # Nombre de retries autorisés
+    retry_delay_ms: int = 2000          # Délai entre retries
+
+    # Vérification
+    success_condition: Optional[SuccessCondition] = None
+
+    # Recovery
+    recovery_action: str = "escape"     # "escape", "undo", "close", "none"
+
+    # Contexte
+    step_id: str = ""                   # Référence vers l'étape WorkflowIR
+    is_optional: bool = False
+
+    def to_dict(self) -> Dict[str, Any]:
+        d = {
+            "node_id": self.node_id,
+            "action_type": self.action_type,
+        }
+        if self.intent:
+            d["intent"] = self.intent
+        if self.strategy_primary:
+            d["strategy_primary"] = self.strategy_primary.to_dict()
+        if self.strategy_fallbacks:
+            d["strategy_fallbacks"] = [s.to_dict() for s in self.strategy_fallbacks]
+        if self.text:
+            d["text"] = self.text
+        if self.keys:
+            d["keys"] = self.keys
+        if self.duration_ms:
+            d["duration_ms"] = self.duration_ms
+        if self.variable_name:
+            d["variable_name"] = self.variable_name
+        d["timeout_ms"] = self.timeout_ms
+        d["max_retries"] = self.max_retries
+        if self.success_condition:
+            d["success_condition"] = self.success_condition.to_dict()
+        d["recovery_action"] = self.recovery_action
+        if self.is_optional:
+            d["is_optional"] = True
+        return d
+
+    @classmethod
+    def from_dict(cls, d: Dict) -> "ExecutionNode":
+        primary = ResolutionStrategy.from_dict(d["strategy_primary"]) if d.get("strategy_primary") else None
+        fallbacks = [ResolutionStrategy.from_dict(f) for f in d.get("strategy_fallbacks", [])]
+        success = SuccessCondition.from_dict(d["success_condition"]) if d.get("success_condition") else None
+        return cls(
+            node_id=d["node_id"],
+            action_type=d["action_type"],
+            intent=d.get("intent", ""),
+            strategy_primary=primary,
+            strategy_fallbacks=fallbacks,
+            text=d.get("text", ""),
+            keys=d.get("keys", []),
+            duration_ms=d.get("duration_ms", 0),
+            variable_name=d.get("variable_name", ""),
+            timeout_ms=d.get("timeout_ms", 10000),
+            max_retries=d.get("max_retries", 1),
+            retry_delay_ms=d.get("retry_delay_ms", 2000),
+            success_condition=success,
+            recovery_action=d.get("recovery_action", "escape"),
+            step_id=d.get("step_id", ""),
+            is_optional=d.get("is_optional", False),
+        )
+
+
+@dataclass
+class ExecutionPlan:
+    """Plan d'exécution versionné — ce que le runtime exécute."""
+    plan_id: str
+    workflow_id: str                    # Référence vers le WorkflowIR source
+    version: int = 1
+    created_at: float = 0.0
+
+    # Nœuds d'exécution (séquence ordonnée)
+    nodes: List[ExecutionNode] = field(default_factory=list)
+
+    # Variables à substituer avant exécution
+    variables: Dict[str, str] = field(default_factory=dict)
+
+    # Configuration globale
+    domain: str = "generic"
+    target_machine: str = ""            # Machine cible
+    target_resolution: str = ""         # "1280x800", "1920x1080"
+
+    # Métriques de compilation
+    total_nodes: int = 0
+    nodes_with_ocr: int = 0             # Résolution OCR (rapide, précis)
+    nodes_with_template: int = 0        # Résolution template (rapide)
+    nodes_with_vlm: int = 0             # Résolution VLM (lent, dernier recours)
+    estimated_duration_s: float = 0.0
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "plan_id": self.plan_id,
+            "workflow_id": self.workflow_id,
+            "version": self.version,
+            "created_at": self.created_at,
+            "domain": self.domain,
+            "target_machine": self.target_machine,
+            "target_resolution": self.target_resolution,
+            "variables": self.variables,
+            "nodes": [n.to_dict() for n in self.nodes],
+            "stats": {
+                "total_nodes": self.total_nodes,
+                "nodes_with_ocr": self.nodes_with_ocr,
+                "nodes_with_template": self.nodes_with_template,
+                "nodes_with_vlm": self.nodes_with_vlm,
+                "estimated_duration_s": round(self.estimated_duration_s, 1),
+            },
+        }
+
+    def to_json(self, indent: int = 2) -> str:
+        return json.dumps(self.to_dict(), ensure_ascii=False, indent=indent)
+
+    @classmethod
+    def from_dict(cls, d: Dict) -> "ExecutionPlan":
+        nodes = [ExecutionNode.from_dict(n) for n in d.get("nodes", [])]
+        stats = d.get("stats", {})
+        return cls(
+            plan_id=d["plan_id"],
+            workflow_id=d.get("workflow_id", ""),
+            version=d.get("version", 1),
+            created_at=d.get("created_at", 0),
+            domain=d.get("domain", "generic"),
+            target_machine=d.get("target_machine", ""),
+            target_resolution=d.get("target_resolution", ""),
+            variables=d.get("variables", {}),
+            nodes=nodes,
+            total_nodes=stats.get("total_nodes", len(nodes)),
+            nodes_with_ocr=stats.get("nodes_with_ocr", 0),
+            nodes_with_template=stats.get("nodes_with_template", 0),
+            nodes_with_vlm=stats.get("nodes_with_vlm", 0),
+            estimated_duration_s=stats.get("estimated_duration_s", 0),
+        )
+
+    @classmethod
+    def from_json(cls, json_str: str) -> "ExecutionPlan":
+        return cls.from_dict(json.loads(json_str))
+
+    def save(self, directory: str) -> Path:
+        dir_path = Path(directory)
+        dir_path.mkdir(parents=True, exist_ok=True)
+        file_path = dir_path / f"{self.plan_id}.json"
+        file_path.write_text(self.to_json(), encoding="utf-8")
+        return file_path
+
+    @classmethod
+    def load(cls, file_path: str) -> "ExecutionPlan":
+        return cls.from_json(Path(file_path).read_text(encoding="utf-8"))
diff --git a/tests/unit/test_execution_compiler.py b/tests/unit/test_execution_compiler.py
new file mode 100644
index 000000000..80c8d81b8
--- /dev/null
+++ b/tests/unit/test_execution_compiler.py
@@ -0,0 +1,264 @@
+"""
+Tests de l'ExecutionCompiler et de l'ExecutionPlan.
+
+Vérifie que :
+- Le compilateur produit un plan déterministe depuis un WorkflowIR
+- Les stratégies de résolution sont correctement compilées (OCR > template > VLM)
+- Les timeouts, retries et recovery sont définis
+- Le plan est sérialisable et versionné
+"""
+
+import json
+import shutil
+import sys
+import tempfile
+from pathlib import Path
+
+import pytest
+
+_ROOT = str(Path(__file__).resolve().parents[2])
+if _ROOT not in sys.path:
+    sys.path.insert(0, _ROOT)
+
+from core.workflow.workflow_ir import WorkflowIR
+from core.workflow.execution_plan import ExecutionPlan, ExecutionNode, ResolutionStrategy, SuccessCondition
+from core.workflow.execution_compiler import ExecutionCompiler
+
+
+# =========================================================================
+# ExecutionPlan — format et sérialisation
+# =========================================================================
+
+
+class TestExecutionPlan:
+
+    def test_serialisation_roundtrip(self):
+        plan = ExecutionPlan(
+            plan_id="plan_test",
+            workflow_id="wf_123",
+            version=1,
+        )
+        plan.nodes.append(ExecutionNode(
+            node_id="n1",
+            action_type="click",
+            intent="Cliquer sur Enregistrer",
+            strategy_primary=ResolutionStrategy(method="ocr", target_text="Enregistrer"),
+            strategy_fallbacks=[ResolutionStrategy(method="vlm", vlm_description="bouton Enregistrer")],
+            success_condition=SuccessCondition(method="title_match", expected_title="Fichier sauvegardé"),
+        ))
+
+        json_str = plan.to_json()
+        plan2 = ExecutionPlan.from_json(json_str)
+
+        assert plan2.plan_id == "plan_test"
+        assert len(plan2.nodes) == 1
+        assert plan2.nodes[0].strategy_primary.method == "ocr"
+        assert len(plan2.nodes[0].strategy_fallbacks) == 1
+        assert plan2.nodes[0].success_condition.method == "title_match"
+
+    def test_save_load(self):
+        tmpdir = tempfile.mkdtemp()
+        try:
+            plan = ExecutionPlan(plan_id="plan_save", workflow_id="wf_1")
+            plan.nodes.append(ExecutionNode(node_id="n1", action_type="click"))
+            path = plan.save(tmpdir)
+
+            plan2 = ExecutionPlan.load(str(path))
+            assert plan2.plan_id == "plan_save"
+            assert len(plan2.nodes) == 1
+        finally:
+            shutil.rmtree(tmpdir)
+
+    def test_node_avec_variable(self):
+        node = ExecutionNode(
+            node_id="n1",
+            action_type="type",
+            text="{patient}",
+            variable_name="patient",
+        )
+        d = node.to_dict()
+        assert d["variable_name"] == "patient"
+        assert d["text"] == "{patient}"
+
+    def test_node_timeout_et_retry(self):
+        node = ExecutionNode(
+            node_id="n1",
+            action_type="click",
+            timeout_ms=5000,
+            max_retries=3,
+            recovery_action="undo",
+        )
+        assert node.timeout_ms == 5000
+        assert node.max_retries == 3
+        assert node.recovery_action == "undo"
+
+
+# =========================================================================
+# ExecutionCompiler — compilation WorkflowIR → ExecutionPlan
+# =========================================================================
+
+
+class TestExecutionCompiler:
+
+    def _make_ir(self):
+        """Créer un WorkflowIR de test."""
+        ir = WorkflowIR.new("Test workflow", domain="generic")
+        ir.add_step(
+            "Ouvrir le fichier",
+            actions=[
+                {"type": "click", "target": "bouton Ouvrir", "anchor_hint": "Ouvrir"},
+                {"type": "wait", "duration_ms": 2000},
+            ],
+            precondition="L'application est ouverte",
+            postcondition="La fenêtre Ouvrir est affichée",
+        )
+        ir.add_step(
+            "Saisir le nom",
+            actions=[
+                {"type": "type", "text": "{nom_fichier}", "variable": True},
+                {"type": "key_combo", "keys": ["enter"]},
+            ],
+        )
+        ir.add_variable("nom_fichier", description="Nom du fichier", default="rapport.pdf")
+        return ir
+
+    def test_compilation_basique(self):
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        assert plan.workflow_id == ir.workflow_id
+        assert plan.total_nodes == 4  # click + wait + type + key_combo
+        assert plan.domain == "generic"
+
+    def test_click_a_strategie_resolution(self):
+        """Un clic doit avoir une stratégie primaire et des fallbacks."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        click_nodes = [n for n in plan.nodes if n.action_type == "click"]
+        assert len(click_nodes) == 1
+
+        click = click_nodes[0]
+        assert click.strategy_primary is not None
+        assert click.strategy_primary.method in ("ocr", "template", "vlm")
+        assert len(click.strategy_fallbacks) >= 1
+
+    def test_ocr_est_prioritaire(self):
+        """Quand du texte est disponible, OCR est la stratégie primaire."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        click = [n for n in plan.nodes if n.action_type == "click"][0]
+        assert click.strategy_primary.method == "ocr"
+        assert click.strategy_primary.target_text == "Ouvrir"
+
+    def test_vlm_est_fallback(self):
+        """Le VLM est toujours en dernier fallback (exception handler)."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        click = [n for n in plan.nodes if n.action_type == "click"][0]
+        vlm_fallbacks = [f for f in click.strategy_fallbacks if f.method == "vlm"]
+        assert len(vlm_fallbacks) >= 1
+
+    def test_type_a_variable(self):
+        """Une action type avec variable a le bon variable_name."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        type_nodes = [n for n in plan.nodes if n.action_type == "type"]
+        assert len(type_nodes) == 1
+        assert type_nodes[0].variable_name == "nom_fichier"
+
+    def test_wait_a_duration(self):
+        """Un wait a la bonne durée."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        wait_nodes = [n for n in plan.nodes if n.action_type == "wait"]
+        assert len(wait_nodes) == 1
+        assert wait_nodes[0].duration_ms == 2000
+
+    def test_click_a_recovery(self):
+        """Un clic a une politique de recovery."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        click = [n for n in plan.nodes if n.action_type == "click"][0]
+        assert click.recovery_action in ("escape", "undo", "close", "none")
+        assert click.max_retries >= 1
+
+    def test_postcondition_devient_success_condition(self):
+        """La postcondition du WorkflowIR devient la condition de succès du nœud."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        click = [n for n in plan.nodes if n.action_type == "click"][0]
+        assert click.success_condition is not None
+        assert "Ouvrir" in click.success_condition.description
+
+    def test_statistiques_compilation(self):
+        """Les statistiques de compilation sont correctes."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        assert plan.total_nodes == 4
+        assert plan.nodes_with_ocr >= 1
+        assert plan.estimated_duration_s > 0
+
+    def test_variables_dans_le_plan(self):
+        """Les variables du WorkflowIR sont dans le plan avec leurs valeurs par défaut."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        assert "nom_fichier" in plan.variables
+        assert plan.variables["nom_fichier"] == "rapport.pdf"
+
+    def test_params_override_defaults(self):
+        """Les params passés au compile écrasent les valeurs par défaut."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir, params={"nom_fichier": "facture_mars.pdf"})
+
+        assert plan.variables["nom_fichier"] == "facture_mars.pdf"
+
+    def test_plan_json_roundtrip(self):
+        """Compiler → JSON → recharger → même plan."""
+        compiler = ExecutionCompiler()
+        ir = self._make_ir()
+        plan = compiler.compile(ir)
+
+        json_str = plan.to_json()
+        plan2 = ExecutionPlan.from_json(json_str)
+
+        assert plan2.total_nodes == plan.total_nodes
+        assert plan2.workflow_id == plan.workflow_id
+        assert len(plan2.nodes) == len(plan.nodes)
+
+    def test_compilation_workflow_vide(self):
+        """Un workflow vide produit un plan vide."""
+        compiler = ExecutionCompiler()
+        ir = WorkflowIR.new("Vide")
+        plan = compiler.compile(ir)
+
+        assert plan.total_nodes == 0
+        assert plan.nodes == []
+
+    def test_plusieurs_domaines(self):
+        """Le compilateur fonctionne pour différents domaines."""
+        compiler = ExecutionCompiler()
+        for domain in ["tim_codage", "comptabilite", "rh_paie", "generic"]:
+            ir = WorkflowIR.new("Test", domain=domain)
+            ir.add_step("Action", actions=[{"type": "click", "target": "bouton"}])
+            plan = compiler.compile(ir)
+            assert plan.domain == domain