feat: graph patch grammar DSL for causal editor

New grammar panel replaces free-text AI prompt with a two-step workflow:
1. NL to Grammar: GPT-4o converts description to patch ops (new endpoint ai-to-grammar)
2. Apply: frontend parser applies grammar locally without AI

Grammar syntax: (+|-|~)(node|edge|coef|input|instruments) key=val key=val
Sign shortcuts: + positive, - negative, = neutral
Auto-positions new nodes when x/y omitted.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

backend/routers/causal_lab.py

@@ -1849,13 +1849,126 @@ def generate_theory(template_id: int):
         raise HTTPException(500, str(e))
 
 
-# ── AI graph modifier ─────────────────────────────────────────────────────────
+# ── AI graph modifier & grammar compiler ─────────────────────────────────────
 
 class AiModifyRequest(BaseModel):
     prompt: str
     current_graph: dict = {}   # client can pass current in-memory state
 
 
+GRAMMAR_SPEC = """\
+Grammar for causal graph patch operations (one operation per line, # = comment):
+
+NODE OPERATIONS
++node id=<id> label="<text>" type=<macro_event|observable|latent|market_asset> [x=<int>] [y=<int>] [instrument=<str>] [formula=<expr>] [unit=<str>]
+~node id=<id> [label="<text>"] [type=...] [x=<int>] [y=<int>] [instrument=<str>]
+-node id=<id>
+
+EDGE OPERATIONS  (sign: + positive  - negative  = neutral)
++edge from=<id> to=<id> [sign=+|-|=] [s=1|2|3] [style=solid|dashed] [label="<text>"] [lag=<int>] [decay=<float|null>]
+~edge from=<id> to=<id> [sign=...] [s=...] [label="<text>"] [lag=<int>] [decay=<float|null>]
+-edge from=<id> to=<id>
+
+COEFFICIENT OPERATIONS
++coef key=<str> value=<float> [desc="<text>"]
+~coef key=<str> [value=<float>] [desc="<text>"]
+-coef key=<str>
+
+INPUT MAPPING
+~input node=<id> source=<surprise|actual_value|user_input|impact_score_scaled> [key=<str>]
+-input node=<id>
+
+GRAPH METADATA
+~instruments EURUSD,SP500,...
+
+Vocabulary:
+- macro_event: source event node (NFP surprise, etc.)
+- observable: directly measurable market variable (OIS rate, US 2Y yield, spread)
+- latent: hidden variable (risk aversion, growth expectations, sentiment)
+- market_asset: traded instrument (EUR/USD, S&P 500 — final output nodes)
+- s=1 weak / s=2 medium (default) / s=3 strong causal link
+- style=dashed for indirect/uncertain channels
+- lag: trading days before effect onset; decay: half-life in days (null = permanent shift)
+"""
+
+
+class AiToGrammarRequest(BaseModel):
+    prompt: str
+    current_graph: dict = {}
+
+
+@router.post("/api/causal-lab/template/{template_id}/ai-to-grammar")
+def ai_to_grammar(template_id: int, body: AiToGrammarRequest):
+    """
+    Convert a natural-language modification request into graph patch grammar.
+    Returns only the grammar lines — user reviews before applying.
+    """
+    try:
+        from services.database import get_conn, get_config
+        from services.causal_graphs import get_template
+        import openai
+
+        key = get_config("openai_api_key") or ""
+        if not key:
+            raise HTTPException(400, "Clé OpenAI manquante dans la configuration")
+
+        conn = get_conn()
+        t = get_template(conn, template_id)
+        conn.close()
+        if not t:
+            raise HTTPException(404, f"Template {template_id} non trouvé")
+
+        gj = body.current_graph if body.current_graph.get("nodes") else t["graph_json"]
+
+        # Compact node/edge summary for context
+        node_summary = "\n".join(
+            f"  {n['id']} ({n['type']}) — {n['label']}"
+            for n in gj.get("nodes", [])
+        )
+        edge_summary = "\n".join(
+            f"  {e['from']} → {e['to']}  sign={e.get('sign','?')} lag={e.get('lag_days','?')} [{e.get('label','')}]"
+            for e in gj.get("edges", [])
+        )
+
+        system_prompt = (
+            "You are a causal graph patch generator for the GeoOptions macro-finance platform.\n"
+            "Convert the user's natural language modification request into graph patch operations.\n"
+            "Return ONLY the grammar lines — no explanation, no markdown, no fences.\n\n"
+            + GRAMMAR_SPEC
+        )
+
+        user_prompt = (
+            f'Current graph nodes:\n{node_summary}\n\n'
+            f'Current graph edges:\n{edge_summary}\n\n'
+            f'Modification request:\n{body.prompt}\n\n'
+            "Return ONLY the grammar patch lines, one per line."
+        )
+
+        client = openai.OpenAI(api_key=key)
+        resp = client.chat.completions.create(
+            model="gpt-4o",
+            messages=[
+                {"role": "system", "content": system_prompt},
+                {"role": "user",   "content": user_prompt},
+            ],
+            temperature=0.15,
+            max_tokens=800,
+        )
+        grammar = (resp.choices[0].message.content or "").strip()
+        # Strip any accidental markdown fences
+        grammar = "\n".join(
+            l for l in grammar.splitlines()
+            if not l.strip().startswith("```")
+        )
+        return {"grammar": grammar}
+
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"[causal_lab] ai_to_grammar {template_id}: {e}")
+        raise HTTPException(500, str(e))
+
+
 @router.post("/api/causal-lab/template/{template_id}/ai-modify")
 def ai_modify_template(template_id: int, body: AiModifyRequest):
     """