Fix expression_compiler: use tier instead of hardcoded model name"

alpha_factory/personas/expression_compiler.py

@@ -42,7 +42,6 @@ ts_decay_linear({{ comps | join(" + ") }}, {{ bp.decay }})
 """,
 }
 
-# Jinja environment
 _env = Environment(loader=BaseLoader())
 
 
@@ -67,11 +66,11 @@ if_else, less, greater, filter, trade_when, ts_backfill, indneutralize
 async def compile_expression(
     blueprint: Blueprint,
     llm: LLMClient,
-    model: str | None = None,
 ) -> Expression:
     """
     Convert a Blueprint to a BRAIN expression.
     Uses Jinja template if archetype is known; LLM fallback for novel structures.
+    Uses tier="tinyfish" — ModelManager resolves to user's selected model.
     """
     # Try template first (95% of cases)
     if blueprint.archetype in TEMPLATES:
@@ -79,7 +78,6 @@ async def compile_expression(
         template = _env.from_string(template_str)
         expr_text = template.render(bp=blueprint).strip()
 
-        # Extract metadata
         fields_used = []
         ops_used = []
         for c in blueprint.components:
@@ -110,7 +108,7 @@ Apply ts_decay_linear with decay={blueprint.decay} if decay > 0."""
     result = await llm.generate_json(
         prompt=user_prompt,
         schema=Expression,
-        model=model or llm.config.tinyfish_model,
+        tier="tinyfish",
         temperature=0.1,
         system_prompt=COMPILER_SYSTEM_PROMPT,
     )
@@ -119,7 +117,6 @@ Apply ts_decay_linear with decay={blueprint.decay} if decay > 0."""
 
 
 def _format_components(bp: Blueprint) -> str:
-    """Format components for the LLM prompt."""
     lines = []
     for i, c in enumerate(bp.components):
         lines.append(