Fix float truncation in arr_get/dict_get arithmetic

The compiler's static type table claimed arr_get / dict_get / arr_min
/ arr_max / arr_sum always return int. They don't — they're
polymorphic over the element type. The lie made

    h sum = arr_get(cur, 1) + rating;     # rating is float

compile to Op::AddInt (the typed fast-path), which calls .to_int() on
both operands and silently truncates the float. Caught by a real
ratings aggregator: summing 4.5 + 3.5 + 5.0 returned 12 instead of
13.0 under VM (tree-walk was fine — its eval_expr inspects values at
runtime, not types at compile time). Tree-walk and VM both wrong-but-
in-different-ways for some shapes.

Removed the polymorphic builtins from the int-return list. They now
have no static return type, so any arithmetic that uses them goes
through Op::Add (the polymorphic dispatch) and respects runtime types.

The four other names that legitimately return int regardless of input
(arr_len, len, arr_index_of, arr_contains) stay in the table.

Found while building examples/recommend/recommend.omc — a real movie
recommendation engine over MovieLens ratings (10k → 100k). VM and
tree-walk now agree on hit counts.

92/92 unit tests pass. 43/43 functional examples produce identical
output under tree-walk and VM.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: RandomCoder-lab

omnimcode-core/src/compiler.rs

@@ -134,11 +134,19 @@ impl Compiler {
                 self.fn_return_types.get(name.as_str()).copied().or_else(|| {
                     // Built-ins whose return type is fixed.
                     match name.as_str() {
+                        // Truly int-returning builtins. Polymorphic ones
+                        // (arr_get, dict_get, arr_min/max/sum — return value
+                        // depends on the element type) are deliberately
+                        // EXCLUDED. Listing them here causes the compiler
+                        // to emit Op::AddInt for `arr_get(...) + x`, which
+                        // calls .to_int() on both operands and silently
+                        // truncates floats. Caught by a real-world float-
+                        // accumulator pattern in examples/recommend.
                         "fibonacci" | "fib" | "is_fibonacci" | "factorial"
                         | "abs" | "floor" | "ceil" | "round" | "is_prime"
                         | "even" | "odd" | "is_even" | "is_odd"
-                        | "len" | "arr_len" | "arr_min" | "arr_max"
-                        | "arr_sum" | "arr_get" | "arr_index_of" | "arr_contains"
+                        | "len" | "arr_len"
+                        | "arr_index_of" | "arr_contains"
                         | "is_singularity" | "resolve_singularity"
                         | "pow_int" | "square" | "cube" | "sign" | "to_int"
                         | "int" | "classify_resonance" | "safe_add" | "safe_sub"