scx_flow: fix autotuner ratio and target_cpu validation

Address two remaining issues from Copilot review PR sched-ext#3525:

1. main.rs: Fix direct_mismatch_ratio denominator
   The original code used cpu_biases as the denominator, but
   direct_local_mismatches can increment even when no bias was
   applied (when no stability bias was available). This could
   cause inflated ratios and incorrect autotuner decisions.

   Now uses cpu_biases.saturating_add(direct_mismatches) as
   the denominator to properly represent all attempts.

2. main.bpf.c: Add target_cpu >= 0 guard
   scx_bpf_task_cpu(p) can return -1 in test scenarios.
   Passing a negative CPU to bpf_cpumask_test_cpu() or using
   SCX_DSQ_LOCAL_ON | target_cpu is undefined behavior.

   Now validates target_cpu before cpumask operations.

Fixes: sched-ext#3525 (comment)
Fixes: sched-ext#3525 (comment)

Author: galpt

scheds/experimental/scx_flow/src/bpf/main.bpf.c

@@ -1362,7 +1362,8 @@ void BPF_STRUCT_OPS(flow_enqueue, struct task_struct *p, u64 enq_flags)
 		if (is_wakeup && !containment_active)
 			enq_flags |= SCX_ENQ_HEAD;
 
-		if (has_wake_target || bpf_cpumask_test_cpu(target_cpu, p->cpus_ptr)) {
+		if (has_wake_target ||
+		    (target_cpu >= 0 && bpf_cpumask_test_cpu(target_cpu, p->cpus_ptr))) {
 			bool should_preempt;
 
 			should_preempt = rt_sensitive_wakeup || ipc_confidence_wakeup ||

scheds/experimental/scx_flow/src/main.rs

@@ -481,7 +481,9 @@ impl AutoTuner {
         let exhaustion_ratio = exhaustions as f64 / positive.max(1) as f64;
         let contained_ratio = contained_enqueues as f64 / positive.max(1) as f64;
         let direct_reject_ratio = direct_rejections as f64 / direct_candidates.max(1) as f64;
-        let direct_mismatch_ratio = direct_mismatches as f64 / cpu_biases.max(1) as f64;
+        let direct_mismatch_attempts = cpu_biases.saturating_add(direct_mismatches);
+        let direct_mismatch_ratio =
+            direct_mismatches as f64 / direct_mismatch_attempts.max(1) as f64;
         let rescue_total = contained_rescues + shared_rescues;
         let rescue_ratio = rescue_total as f64 / dispatch_total.max(1) as f64;
         let latency_dispatch_ratio = total_latency_dispatches as f64 / dispatch_total.max(1) as f64;