fix: remove determinism diagnostics, restore IQL step

Training is CONFIRMED BIT-IDENTICAL across runs (all per-step weight
checksums match). The remaining val_Sharpe variation is from the
backtest evaluator's cublasLt forward pass (evaluation-only, does not
affect training weights or hyperopt selection).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml/src/trainers/dqn/fused_training.rs

@@ -1995,6 +1995,22 @@ impl FusedTrainingCtx {
         self.trainer.params().raw_ptr()
     }
 
+    /// Compute a simple checksum of the flat params buffer (GPU→CPU readback).
+    /// For determinism diagnostics only — synchronizes the stream.
+    pub(crate) fn params_checksum(&self) -> anyhow::Result<u64> {
+        let n = self.trainer.total_params();
+        let params = self.trainer.params();
+        let host: Vec<f32> = self.stream.clone_dtoh(params)
+            .map_err(|e| anyhow::anyhow!("params DtoH for checksum: {e}"))?;
+        let mut hash: u64 = 0xcbf29ce484222325; // FNV-1a
+        for &v in &host[..n] {
+            let bits = v.to_bits();
+            hash ^= bits as u64;
+            hash = hash.wrapping_mul(0x100000001b3);
+        }
+        Ok(hash)
+    }
+
     pub(crate) fn trainer_v_range_buf_ptr(&self) -> u64 { self.trainer.v_range_buf_ptr() }
     pub(crate) fn adapt_v_range(&mut self, q_mean: f32, q_variance: f32) -> bool { self.trainer.adapt_v_range(q_mean, q_variance) }
     pub(crate) fn adapt_v_range_with_gap(&mut self, q_mean: f32, q_variance: f32, q_gap: f32) -> bool { self.trainer.adapt_v_range_with_gap(q_mean, q_variance, q_gap) }